<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: James Ngandu</title>
    <description>The latest articles on DEV Community by James Ngandu (@mukuvi).</description>
    <link>https://dev.to/mukuvi</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F1177239%2Fccb72090-3f30-44df-b801-8aec157f5609.jpg</url>
      <title>DEV Community: James Ngandu</title>
      <link>https://dev.to/mukuvi</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/mukuvi"/>
    <language>en</language>
    <item>
      <title>From DevFest Rift Valley to DevFest Nairobi 2026: Why I Keep Coming Back</title>
      <dc:creator>James Ngandu</dc:creator>
      <pubDate>Sun, 15 Nov 2026 14:20:21 +0000</pubDate>
      <link>https://dev.to/mukuvi/from-devfest-rift-valley-to-devfest-nairobi-2026-why-i-keep-coming-back-577l</link>
      <guid>https://dev.to/mukuvi/from-devfest-rift-valley-to-devfest-nairobi-2026-why-i-keep-coming-back-577l</guid>
      <description>&lt;p&gt;Some tech events leave you with more than just notes and photos. They introduce you to new ideas, people building interesting things, and perspectives that make you rethink how you approach technology.&lt;/p&gt;

&lt;p&gt;For me, DevFest has become one of those experiences.&lt;/p&gt;

&lt;p&gt;My journey with DevFest started at DevFest Rift Valley. This year, I had the opportunity to experience DevFest Nairobi 2026, hosted by Google Developer Groups Nairobi at the Sarit Expo Centre.&lt;/p&gt;

&lt;p&gt;Under the theme &lt;strong&gt;Build. Secure. Scale.&lt;/strong&gt;, the event brought together developers, engineers, founders, designers, technology enthusiasts and people passionate about building the future.&lt;/p&gt;

&lt;p&gt;And there was quite a lot to explore.&lt;/p&gt;

&lt;h2&gt;
  
  
  Beyond the usual conversations about AI
&lt;/h2&gt;

&lt;p&gt;AI and agentic applications were among the major themes throughout the event. There were discussions around building production agents with Gemini, solving complex problems with AI, and moving beyond traditional chatbots towards more capable AI-powered products.&lt;/p&gt;

&lt;p&gt;What I found interesting was the shift in conversation. We are moving from simply experimenting with AI APIs to thinking about how to build, deploy, secure and maintain AI systems that solve actual problems.&lt;/p&gt;

&lt;p&gt;There were also sessions covering cloud computing, cybersecurity, APIs, IoT, mobile development, accessibility and software engineering.&lt;/p&gt;

&lt;p&gt;Some topics that particularly caught my attention included securing autonomous agents in production, designing APIs for AI agents, building high-throughput and low-latency IoT systems with Go, and architecting secure SaaS applications on Google Cloud.&lt;/p&gt;

&lt;p&gt;These are areas that connect closely with the engineering challenges developers face when building real-world systems.&lt;/p&gt;

&lt;h2&gt;
  
  
  The value of being part of a developer community
&lt;/h2&gt;

&lt;p&gt;One thing I appreciate about DevFest is that the experience extends beyond the stage.&lt;/p&gt;

&lt;p&gt;Meeting other developers, exchanging ideas, discovering what people are building and having conversations about the industry are equally important.&lt;/p&gt;

&lt;p&gt;Technology evolves quickly, and sometimes the most valuable thing you can take away from an event is a conversation with someone approaching a problem differently.&lt;/p&gt;

&lt;p&gt;It is also encouraging to see the developer ecosystem in Nairobi bringing together people with different skills, experiences and ambitions.&lt;/p&gt;

&lt;h2&gt;
  
  
  A journey that keeps growing
&lt;/h2&gt;

&lt;p&gt;I also want to appreciate &lt;strong&gt;Wayne Jotham Gakuo&lt;/strong&gt;. His enthusiasm and influence around DevFest and the Google developer ecosystem have contributed to my growing interest in these events and Google's technologies over the years.&lt;/p&gt;

&lt;p&gt;From my experience at DevFest Rift Valley to now attending DevFest Nairobi, that interest has continued to grow.&lt;/p&gt;

&lt;p&gt;I have developed a greater appreciation for the work Google Developer Groups communities do in creating spaces where developers can learn, collaborate and share knowledge.&lt;/p&gt;

&lt;h2&gt;
  
  
  Looking ahead
&lt;/h2&gt;

&lt;p&gt;DevFest Nairobi 2026 was a reminder of how much there is still to learn and build.&lt;/p&gt;

&lt;p&gt;Whether it is AI, cloud infrastructure, cybersecurity or software engineering, the opportunities are enormous. But so are the responsibilities that come with building technology that people and businesses depend on.&lt;/p&gt;

&lt;p&gt;I left with new ideas, valuable connections and more areas I want to explore.&lt;/p&gt;

&lt;p&gt;Thank you to GDG Nairobi, the speakers, organizers, volunteers and everyone who contributed to making the event happen.&lt;/p&gt;

&lt;p&gt;Until the next DevFest.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Build. Secure. Scale.&lt;/strong&gt;&lt;/p&gt;




&lt;p&gt;&lt;em&gt;What I keep taking away from these events is simple: the best way to understand where technology is going is to stay curious, keep building and remain connected to the people building it.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>security</category>
      <category>agents</category>
      <category>android</category>
    </item>
    <item>
      <title>Letting users bring their own AI key without storing it in plaintext</title>
      <dc:creator>James Ngandu</dc:creator>
      <pubDate>Tue, 10 Nov 2026 17:14:08 +0000</pubDate>
      <link>https://dev.to/mukuvi/letting-users-bring-their-own-ai-key-without-storing-it-in-plaintext-2jfi</link>
      <guid>https://dev.to/mukuvi/letting-users-bring-their-own-ai-key-without-storing-it-in-plaintext-2jfi</guid>
      <description>&lt;p&gt;A single AI key shared by every user of a small app is a bad deal. The free tier runs out, one noisy user consumes the quota, and the owner pays for everyone. Letting each reader bring their own key fixes that, but it creates a different problem: now the app stores secrets that belong to other people.&lt;/p&gt;

&lt;p&gt;ARN.IO is a reading app with a book assistant. Readers can ask about a summary, a chapter or a theme. I added a bring-your-own-key setting this week. The rule is simple: a key can go in, but it never comes back out.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the user sees
&lt;/h2&gt;

&lt;p&gt;The account page has an AI tab with four fields:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Provider&lt;/li&gt;
&lt;li&gt;Model&lt;/li&gt;
&lt;li&gt;API base URL&lt;/li&gt;
&lt;li&gt;API key&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The provider list covers Groq, OpenAI, OpenRouter, DeepSeek, Google Gemini, Mistral and a custom OpenAI-compatible endpoint. Choosing a preset fills in the base URL and a default model. The key field is a password input and is never prefilled, even when a key is already saved.&lt;/p&gt;

&lt;p&gt;After saving, the page shows which provider is active and offers a Remove Key button. The settings endpoint returns &lt;code&gt;hasKey: true&lt;/code&gt; and nothing that can be used to recover the secret.&lt;/p&gt;

&lt;h2&gt;
  
  
  The key is write-only
&lt;/h2&gt;

&lt;p&gt;The database table is deliberately small:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight sql"&gt;&lt;code&gt;&lt;span class="k"&gt;CREATE&lt;/span&gt; &lt;span class="k"&gt;TABLE&lt;/span&gt; &lt;span class="k"&gt;IF&lt;/span&gt; &lt;span class="k"&gt;NOT&lt;/span&gt; &lt;span class="k"&gt;EXISTS&lt;/span&gt; &lt;span class="n"&gt;user_ai_settings&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;
  &lt;span class="n"&gt;user_id&lt;/span&gt; &lt;span class="kt"&gt;INTEGER&lt;/span&gt; &lt;span class="k"&gt;PRIMARY&lt;/span&gt; &lt;span class="k"&gt;KEY&lt;/span&gt; &lt;span class="k"&gt;REFERENCES&lt;/span&gt; &lt;span class="nf"&gt;users&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;id&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;ON&lt;/span&gt; &lt;span class="k"&gt;DELETE&lt;/span&gt; &lt;span class="k"&gt;CASCADE&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;provider&lt;/span&gt; &lt;span class="kt"&gt;TEXT&lt;/span&gt; &lt;span class="k"&gt;NOT&lt;/span&gt; &lt;span class="kc"&gt;NULL&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;base_url&lt;/span&gt; &lt;span class="kt"&gt;TEXT&lt;/span&gt; &lt;span class="k"&gt;NOT&lt;/span&gt; &lt;span class="kc"&gt;NULL&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;model&lt;/span&gt; &lt;span class="kt"&gt;TEXT&lt;/span&gt; &lt;span class="k"&gt;NOT&lt;/span&gt; &lt;span class="kc"&gt;NULL&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;api_key_encrypted&lt;/span&gt; &lt;span class="kt"&gt;TEXT&lt;/span&gt; &lt;span class="k"&gt;NOT&lt;/span&gt; &lt;span class="kc"&gt;NULL&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;created_at&lt;/span&gt; &lt;span class="kt"&gt;TIMESTAMPTZ&lt;/span&gt; &lt;span class="k"&gt;DEFAULT&lt;/span&gt; &lt;span class="nb"&gt;NOW&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;updated_at&lt;/span&gt; &lt;span class="kt"&gt;TIMESTAMPTZ&lt;/span&gt; &lt;span class="k"&gt;DEFAULT&lt;/span&gt; &lt;span class="nb"&gt;NOW&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The key is encrypted with AES-256-GCM before it reaches the row. The encryption key lives in an environment variable, not in PostgreSQL. The stored value is the initialization vector, the authentication tag and the ciphertext, joined with dots:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="k"&gt;export&lt;/span&gt; &lt;span class="kd"&gt;function&lt;/span&gt; &lt;span class="nf"&gt;encryptSecret&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;value&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;iv&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;crypto&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;randomBytes&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;12&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;cipher&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;crypto&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;createCipheriv&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"aes-256-gcm"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nf"&gt;getKey&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;iv&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;encrypted&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;Buffer&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;concat&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;
    &lt;span class="n"&gt;cipher&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;update&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nb"&gt;String&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;value&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"utf8"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;cipher&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;final&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;tag&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;cipher&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;getAuthTag&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;
    &lt;span class="n"&gt;iv&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;toString&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"base64"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;tag&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;toString&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"base64"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;encrypted&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;toString&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"base64"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;join&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"."&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;GCM was chosen because it gives confidentiality and integrity in one step. A modified ciphertext fails to decrypt instead of returning garbage. The key is 32 bytes and is validated at startup. If the environment value is missing or the wrong length, saving a key fails with a configuration error instead of writing something unreadable.&lt;/p&gt;

&lt;p&gt;There is one operational consequence. If the encryption key is rotated, every saved user key becomes unreadable. The app treats that as a failed key, falls back to the built-in assistant, and the user has to enter the key again. In a larger system I would store a key version alongside the ciphertext so old keys can be decrypted during a rotation window.&lt;/p&gt;

&lt;h2&gt;
  
  
  The API is intentionally boring
&lt;/h2&gt;

&lt;p&gt;The settings endpoint returns only the configuration a browser needs to render the form:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="nl"&gt;"settings"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="nl"&gt;"provider"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"groq"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="nl"&gt;"baseUrl"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"https://api.groq.com/openai/v1"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="nl"&gt;"model"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"llama-3.3-70b-versatile"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="nl"&gt;"hasKey"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;There is no endpoint that returns the key. The only time the server decrypts it is immediately before a provider request, and the decrypted value stays in memory for that request.&lt;/p&gt;

&lt;p&gt;Saving uses &lt;code&gt;PUT /api/auth/ai-settings&lt;/code&gt;. If the user leaves the key blank for an existing setup, the server keeps the stored ciphertext. That makes it possible to change the model without asking for the secret again.&lt;/p&gt;

&lt;p&gt;Changing the provider with a blank key is rejected. A Groq key is not an OpenAI key, and silently pairing the old secret with a new service would only produce confusing failures.&lt;/p&gt;

&lt;h2&gt;
  
  
  Provider presets, not provider lock-in
&lt;/h2&gt;

&lt;p&gt;The server stores three values per user: the provider id, the base URL and the model. The provider id is mostly a label and a source of defaults. The request path is the same for all of them because every preset uses an OpenAI-compatible chat completions endpoint.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Provider&lt;/th&gt;
&lt;th&gt;Base URL&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Groq&lt;/td&gt;
&lt;td&gt;&lt;code&gt;https://api.groq.com/openai/v1&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;OpenAI&lt;/td&gt;
&lt;td&gt;&lt;code&gt;https://api.openai.com/v1&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;OpenRouter&lt;/td&gt;
&lt;td&gt;&lt;code&gt;https://openrouter.ai/api/v1&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;DeepSeek&lt;/td&gt;
&lt;td&gt;&lt;code&gt;https://api.deepseek.com/v1&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Google Gemini&lt;/td&gt;
&lt;td&gt;&lt;code&gt;https://generativelanguage.googleapis.com/v1beta/openai&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Mistral&lt;/td&gt;
&lt;td&gt;&lt;code&gt;https://api.mistral.ai/v1&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The custom option accepts any public HTTPS endpoint with the same shape. That keeps the door open for providers I have not listed without turning the server into a provider-specific integration for each one.&lt;/p&gt;

&lt;h2&gt;
  
  
  A custom base URL is an SSRF risk
&lt;/h2&gt;

&lt;p&gt;This was the part that needed the most care. A user-supplied base URL means the server will make an outbound request to an address chosen by that user. Without checks, someone could point it at &lt;code&gt;http://127.0.0.1:5432&lt;/code&gt; or a cloud metadata service.&lt;/p&gt;

&lt;p&gt;The validator does five things before a custom URL is saved:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;It requires HTTPS.&lt;/li&gt;
&lt;li&gt;It rejects URLs with embedded credentials, query strings or fragments.&lt;/li&gt;
&lt;li&gt;It resolves the hostname and checks every returned address.&lt;/li&gt;
&lt;li&gt;It blocks private IPv4 ranges, private IPv6 ranges, localhost, &lt;code&gt;.local&lt;/code&gt; names and metadata hosts.&lt;/li&gt;
&lt;li&gt;It rejects the request if the hostname cannot be resolved.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The provider request also uses &lt;code&gt;redirect: "error"&lt;/code&gt; and a timeout. A redirect to an internal address is not followed.&lt;/p&gt;

&lt;p&gt;This is not a complete SSRF defense. A DNS rebinding attack can resolve to a public address during validation and a private address during the request. A stricter design would resolve once, pin the address, and connect to that exact address while keeping the host header. For a small reading app, the validation plus no redirects plus a short timeout is a reasonable barrier. I would not treat it as sufficient for a high-risk multi-tenant platform without the address pinning step.&lt;/p&gt;

&lt;h2&gt;
  
  
  How a request chooses a key
&lt;/h2&gt;

&lt;p&gt;The chat path has four possible outcomes:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;The user's saved key is decrypted and used.&lt;/li&gt;
&lt;li&gt;If the user has no key, the server's own key is used.&lt;/li&gt;
&lt;li&gt;If the user's key fails and a server key exists, the server key is tried.&lt;/li&gt;
&lt;li&gt;If neither works, the built-in assistant answers.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The built-in fallback is not a language model. It handles summaries from the book description, chapter lists, author questions and a short excerpt from the current chapter. That keeps the assistant usable when a provider is down or a key has expired.&lt;/p&gt;

&lt;p&gt;Key errors are logged on the server, not shown in the chat. A rejected key should not turn the reading experience into a configuration lesson. The user sees a normal answer and can fix the key in the account page when they choose to.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I would change next
&lt;/h2&gt;

&lt;p&gt;The current version is a good first pass, but there are four things I would add before calling it finished:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;A test button.&lt;/strong&gt; Save the key, then send one small request and report whether the provider accepted it. Right now the first real chat is the test.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Per-user rate limits.&lt;/strong&gt; A user key protects the owner's quota, but it does not protect the app from a user who sends hundreds of requests in a minute.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Key versioning.&lt;/strong&gt; Store an encryption key version so a rotation does not invalidate every saved key at once.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A managed secret store.&lt;/strong&gt; Application-level AES-GCM is fine for a small deployment, but a dedicated secret manager reduces the number of places the master key can appear.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The principle I keep coming back to is simple. A user's API key is a loan, not a copy. The app should use it for the request in front of it, store only what it must, and make removal easy.&lt;/p&gt;

&lt;p&gt;ARN.IO is live at &lt;a href="https://arnio.mukuvi.tech" rel="noopener noreferrer"&gt;https://arnio.mukuvi.tech&lt;/a&gt;. The AI tab is in the profile page.&lt;/p&gt;

&lt;h2&gt;
  
  
  Related reading
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://dev.to/c9dn/how-to-let-users-bring-their-own-openai-or-anthropic-api-keys-without-storing-them-in-plaintext-12m"&gt;How to let users bring their own OpenAI or Anthropic API keys (without storing them in plaintext)&lt;/a&gt; by Shlok Madhekar&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://dev.to/fuadhusnan_f44f3e13/how-to-safely-store-and-encrypt-api-credentials-in-production-e72"&gt;How to Safely Store and Encrypt API Credentials in Production&lt;/a&gt; by Fuad Husnan&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://dev.to/craftorytool/my-ssrf-guard-blocked-127001-and-almost-nothing-else-15l4"&gt;My SSRF guard blocked 127.0.0.1 and almost nothing else&lt;/a&gt; by Anirban Sarkar&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>security</category>
      <category>node</category>
      <category>webdev</category>
      <category>postgres</category>
    </item>
    <item>
      <title>Authorization before the scanner: building King Malhare</title>
      <dc:creator>James Ngandu</dc:creator>
      <pubDate>Mon, 09 Nov 2026 16:46:18 +0000</pubDate>
      <link>https://dev.to/mukuvi/authorization-before-the-scanner-building-king-malhare-5fgo</link>
      <guid>https://dev.to/mukuvi/authorization-before-the-scanner-building-king-malhare-5fgo</guid>
      <description>&lt;p&gt;Most vulnerability scanners are easy to point at a host. You give them a URL, choose a profile, and they start sending traffic. The hard part is proving that the host was allowed to be scanned in the first place.&lt;/p&gt;

&lt;p&gt;King Malhare is a self-hosted security assessment platform I built around one rule: no active scan runs without a valid authorization record, and that record has to survive the trip from the API to the scanner worker. The live instance is at &lt;a href="https://kingmalhare.mukuvi.tech" rel="noopener noreferrer"&gt;https://kingmalhare.mukuvi.tech&lt;/a&gt;. The code is proprietary, so this post focuses on the architecture and the decisions behind it.&lt;/p&gt;

&lt;h2&gt;
  
  
  The stack
&lt;/h2&gt;

&lt;p&gt;The platform runs as four services:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;A Go API built on chi and pgx&lt;/li&gt;
&lt;li&gt;PostgreSQL 16 as the source of truth&lt;/li&gt;
&lt;li&gt;Redis 7 for the job queue, cancellation and progress fan-out&lt;/li&gt;
&lt;li&gt;Python workers that run the scanner adapters&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The console is React with Vite and TanStack Query. Caddy handles TLS in the production Compose stack.&lt;/p&gt;

&lt;p&gt;I split the API from the workers for a specific reason. Scanners deal with hostile input in both directions. They parse output from arbitrary targets, and some of them wrap binaries that can be tricked. I did not want any of that inside the public API process.&lt;/p&gt;

&lt;p&gt;The worker has no database URL, no published port, and no route from the internet. It runs on an internal Docker network with &lt;code&gt;cap_drop: ALL&lt;/code&gt;, &lt;code&gt;no-new-privileges&lt;/code&gt;, a read-only root filesystem, a writable tmpfs for &lt;code&gt;/tmp&lt;/code&gt;, and CPU, memory and PID limits. It talks to the API with a service token. The API is the only service that can read or write the database.&lt;/p&gt;

&lt;h2&gt;
  
  
  Authorization is enforced in three places
&lt;/h2&gt;

&lt;p&gt;A common mistake in security tooling is to check permission in the UI and trust the result. A button that is hidden is not an authorization control. King Malhare checks scope at three points.&lt;/p&gt;

&lt;p&gt;First, the API checks before it queues anything. The asset must belong to the organization. The authorization must belong to the asset, be active and not expired. If the scanner or profile is intrusive, the authorization must explicitly allow intrusive testing. The requested parameters are checked against an allowlist. The target is resolved without scanning and rejected if it points at a metadata address or a private range that the authorization does not cover.&lt;/p&gt;

&lt;p&gt;Second, the worker checks again before it runs a binary. Scope is copied into the job envelope. The worker rechecks the expiry, the intrusive flag and the approved target host. It resolves DNS again and checks every returned address. Link-local &lt;code&gt;169.254.0.0/16&lt;/code&gt; and metadata hostnames are always rejected. Private addresses are allowed only when the authorization says so and the asset is marked as a lab or internal environment.&lt;/p&gt;

&lt;p&gt;Third, the scanner itself runs with a fixed argument vector. There is no shell. User supplied strings never become a shell token. The worker uses &lt;code&gt;exec&lt;/code&gt;, a per-job timeout, a maximum captured output size, and a clean environment.&lt;/p&gt;

&lt;p&gt;The job envelope is small on purpose:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"scan_id"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"7f2c..."&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"org_id"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"1a9b..."&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"scanner"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"nuclei"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"target"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"https://app.example.com"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"scope"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"authorization_id"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"d4e1..."&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"allow_private"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"intrusive_allowed"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"expires_at"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"2026-12-01T00:00:00Z"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"approved_targets"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"app.example.com"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Every field that matters is copied into the job instead of being looked up later from a mutable row. If someone edits the authorization after the job is queued, the job still carries the scope that was valid when it was created.&lt;/p&gt;

&lt;h2&gt;
  
  
  Findings are normalized and deduplicated
&lt;/h2&gt;

&lt;p&gt;Each scanner produces its own format. Nmap gives ports and services. Nuclei gives JSONL. ZAP gives alerts. The workers map all of it into one result shape before it reaches the database.&lt;/p&gt;

&lt;p&gt;The important field is the fingerprint:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;sha256(scanner | normalized title | endpoint | cve-or-cwe)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That fingerprint lets a re-run update &lt;code&gt;last_seen&lt;/code&gt; without creating a duplicate. &lt;code&gt;first_seen&lt;/code&gt; stays intact, and the triage status survives. A finding that was accepted as a false positive stays accepted after the next scan.&lt;/p&gt;

&lt;p&gt;The adapters are deliberately conservative. A missing security header becomes a finding only when the header is actually absent. A TLS issue becomes a finding only when the handshake proves the condition. A path is reported only when the response is not a 404. If a binary is not installed, the adapter returns &lt;code&gt;unavailable&lt;/code&gt; with the path it checked. It does not invent a result to fill the dashboard.&lt;/p&gt;

&lt;p&gt;That last rule matters more than it sounds. A security tool that reports fiction is worse than a tool that reports nothing, because it trains the operator to stop checking.&lt;/p&gt;

&lt;h2&gt;
  
  
  Queue, progress and cancellation
&lt;/h2&gt;

&lt;p&gt;Redis holds the job queue. The API pushes a JSON job and a worker pops it. Before running anything, the worker claims the job through the API, and the API performs a conditional update from &lt;code&gt;queued&lt;/code&gt; to &lt;code&gt;running&lt;/code&gt;. Two workers cannot claim the same job.&lt;/p&gt;

&lt;p&gt;Progress is sent back through the API and published on Redis. The browser receives it over Server Sent Events. The UI also polls every 15 seconds as a fallback, because SSE connections drop and a progress bar that freezes is not useful.&lt;/p&gt;

&lt;p&gt;Cancellation has two levels. A single scan sets a cancel key and the worker checks it between phases. The organization level emergency stop sets a kill flag and marks every queued or running job cancelled before a worker picks up more work.&lt;/p&gt;

&lt;p&gt;The API also has a stale job reaper. If a worker dies mid-job, the scan is marked failed with an actionable message. It is not automatically re-queued. Re-running an intrusive scan because a worker crashed is not a safe default.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I would do differently
&lt;/h2&gt;

&lt;p&gt;The scanner integration was not the hard part. The hard part was making the authorization trail, the evidence and the report line up.&lt;/p&gt;

&lt;p&gt;A few lessons from the build:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Put scope enforcement in the API and repeat it in the worker.&lt;/strong&gt; Do not make the worker trust the queue. The queue is a transport, not a policy engine.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Make adapters conservative.&lt;/strong&gt; A smaller number of findings that an operator trusts is better than a large number that has to be manually filtered. False positives cost more than missed checks in a tool like this.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Do not add a queue unless you need one.&lt;/strong&gt; Cameron Bennett wrote about reaching for Postgres before a queue, and it made me recheck the decision. I kept Redis because I needed multiple workers, per-scan cancellation and progress fan-out. If the requirement had only been "run this later", Postgres would have been enough.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Tenant isolation belongs in the query layer.&lt;/strong&gt; Saint Zero Day's post on multi-tenant self-hosted security platforms covers this well. Hiding another organization's asset in the UI does nothing if the query still returns it. Every business query in King Malhare is scoped by &lt;code&gt;org_id&lt;/code&gt;, and the tests include cross-tenant reads and writes.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The report is part of the product.&lt;/strong&gt; A finding without evidence, impact and remediation guidance is just a line in a table. The report engine embeds the same data the console shows, including screenshots where the scanner captured them.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is next
&lt;/h2&gt;

&lt;p&gt;The current native cloud posture checks cover AWS. Azure and GCP have the adapter path and documented CLI wrappers, but native SDK checks are still on the list. I also want to keep improving the report templates and add more lab fixtures so adapter changes are tested against known conditions rather than hand written expectations.&lt;/p&gt;

&lt;p&gt;King Malhare is live at &lt;a href="https://kingmalhare.mukuvi.tech" rel="noopener noreferrer"&gt;https://kingmalhare.mukuvi.tech&lt;/a&gt;. The scanner is the easy part. The authorization trail, the evidence and the report are what make the result useful.&lt;/p&gt;

&lt;h2&gt;
  
  
  Related reading
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://dev.to/matarturo/more-than-just-a-scanner-architecture-of-a-self-hosted-dast-engine-for-spas-and-cicd-2m3j"&gt;More Than Just a Scanner: Architecture of a Self-Hosted DAST Engine for SPAs and CI/CD&lt;/a&gt; by Arturo Enrique Mata Garcia&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://dev.to/saint_zero_day/one-install-many-customers-building-airtight-multi-tenancy-into-a-self-hosted-security-platform-kl2"&gt;One install, many customers: building airtight multi-tenancy into a self-hosted security platform&lt;/a&gt; by Saint Zero Day&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://dev.to/camerondev/go-postgres-and-the-queue-i-didnt-need-3f2g"&gt;Go, Postgres, and the queue I didn't need&lt;/a&gt; by Cameron Bennett&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>security</category>
      <category>go</category>
      <category>python</category>
      <category>postgres</category>
    </item>
    <item>
      <title>Every first pull request deserves a doorman</title>
      <dc:creator>James Ngandu</dc:creator>
      <pubDate>Fri, 06 Nov 2026 07:06:42 +0000</pubDate>
      <link>https://dev.to/mukuvi/every-first-pull-request-deserves-a-doorman-14j0</link>
      <guid>https://dev.to/mukuvi/every-first-pull-request-deserves-a-doorman-14j0</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/mlh-hackathon"&gt;MLH x DEV Writing Challenge&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;The most fragile moment in open source is not a merge conflict. It is the twenty minutes after someone opens their first pull request.&lt;/p&gt;

&lt;p&gt;They have read the code, worked up the nerve, forked, branched, pushed, and written the thing that always feels too short. Then they wait. What usually comes back is silence, because maintainers are busy, or a two-line review that reads like a door closing.&lt;/p&gt;

&lt;p&gt;Doorman is a GitHub Action that stands at that door.&lt;/p&gt;

&lt;p&gt;When a first-time contributor opens a pull request, Doorman reads it and leaves one comment in about a second. It says what is already good, then lists the few things a maintainer would have asked for anyway:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;a title that says what changed&lt;/li&gt;
&lt;li&gt;a description that explains why&lt;/li&gt;
&lt;li&gt;an issue linked&lt;/li&gt;
&lt;li&gt;a branch name someone can read&lt;/li&gt;
&lt;li&gt;commit messages that mean something&lt;/li&gt;
&lt;li&gt;a change small enough to review&lt;/li&gt;
&lt;li&gt;tests, when code changed&lt;/li&gt;
&lt;li&gt;no binary files that should not be there&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Eight checks, no model, no call to a language model API, and nothing outside the Python standard library.&lt;/p&gt;

&lt;p&gt;The rules are the point. A contributor's first interaction with a project should not depend on a model's mood. A check that says "your description is 10 characters, add what was wrong and what you changed" is boring, and boring is what trust is made of. You can read every line that decides what the bot says to you.&lt;/p&gt;

&lt;p&gt;Three rules shaped the writing more than the code did.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Good news first.&lt;/strong&gt; Every comment starts with what already works. There is always something.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;No review language.&lt;/strong&gt; No "you must", no "changes requested". A finding is either "needs fixing" or "optional", and the comment says plainly that none of it is a rejection.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;A person is still coming.&lt;/strong&gt; The last line of every comment says a human will read it. The action is a doorman, not a gatekeeper.&lt;/p&gt;

&lt;p&gt;There is also a hidden marker in the comment, so if Doorman has already greeted a pull request, a new commit will not set off a second greeting.&lt;/p&gt;

&lt;p&gt;The whole thing is about 600 lines of standard library Python with 28 tests, and it runs on &lt;code&gt;pull_request_target&lt;/code&gt; with only the permissions it needs. It never checks out or runs code from the fork, which matters when the entire job is to comment on a stranger's branch.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;Here is the real output on a first pull request that is a little rough, the kind that arrives at every project:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Hi @newcontributor, welcome to your first pull request here. Thank you for taking the time.

**4 things worth a look, 4 already good.**

### Already good
- **Commit messages describe the change.**
- **Focused change: 1 file.**
- **No code changed, so no test was needed.**
- **No unexpected binary files.**

### Worth a look
- **The title "Update README.md" does not say what changed.** Vague titles get lost in a busy pull request list.
  - Try something like "Fix KeyError when a user has no profile".
  - _Needs fixing_
- **The description is very short.** 10 characters.
  - Add what was wrong, what you changed, and how you checked it.
  - _Needs fixing_
- **No issue linked.** Maintainers use the issue to remember why the change exists.
  - If this closes one, add a line like "Closes #42" to the description.
  - _Optional_

None of this is a rejection. Push a new commit to the same branch and everything updates, including this note.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;On a pull request that already does the basics, it says one line and gets out of the way: &lt;strong&gt;Everything I check looks good.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;You can run it yourself with no token, against the two recorded payloads in the repo:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;python3 &lt;span class="nt"&gt;-m&lt;/span&gt; doorman review &lt;span class="nt"&gt;--event&lt;/span&gt; examples/first_pull_request.json
python3 &lt;span class="nt"&gt;-m&lt;/span&gt; doorman review &lt;span class="nt"&gt;--event&lt;/span&gt; examples/steady_pull_request.json
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Repo: &lt;a href="https://github.com/mukuvi/doorman" rel="noopener noreferrer"&gt;github.com/mukuvi/doorman&lt;/a&gt;. MIT licensed, tagged v0.1.0. The cover image above is drawn by a script in the repo, because I could not bring myself to pay for a design tool for a weekend project.&lt;/p&gt;

&lt;h2&gt;
  
  
  Partner Technologies
&lt;/h2&gt;

&lt;p&gt;The partner technology here is GitHub itself: Actions, the REST API, &lt;code&gt;pull_request_target&lt;/code&gt;, the token model, and the label system.&lt;/p&gt;

&lt;p&gt;Most of the interesting work was in the permissions. A pull request from a fork gets a read-only token by default, which is exactly right, and also means a plain &lt;code&gt;pull_request&lt;/code&gt; workflow cannot leave a comment. &lt;code&gt;pull_request_target&lt;/code&gt; fixes that, but it runs in the context of the base repository, so checking out the fork becomes a security decision. Doorman never checks out the fork. It reads the event payload, calls the API for the changed files and commits, and posts a comment. Two write permissions, and no code execution.&lt;/p&gt;

&lt;p&gt;I also used the label API to add a &lt;code&gt;first-time-contributor&lt;/code&gt; label, so a maintainer can filter for the exact people who need a slower, kinder review.&lt;/p&gt;

&lt;p&gt;I did not use Auth0 or Tiger Data this time, even though both were on the week's schedule. Doorman has no users to authenticate and no rows to store. Adding either would have been decoration.&lt;/p&gt;

&lt;h2&gt;
  
  
  Hackathon Experience
&lt;/h2&gt;

&lt;p&gt;I built this during Global Hack Week: Builders Week, the MLH week that runs November 6 to 12.&lt;/p&gt;

&lt;p&gt;The theme is building, which sounds obvious until you notice how much of a hackathon is usually setup. Accounts, keys, a database you do not need, an auth flow for a demo with no users. I wanted the opposite of that: one repository, one action, one file you can read in an evening.&lt;/p&gt;

&lt;p&gt;Building from Nairobi means the livestreams land late. Most of my work happened after eleven at night, with the workshop recordings open in one window and a terminal in the other. I had the GitHub Copilot session running while I wrote the checks, which is a strange way to spend an evening about automation: writing careful deterministic rules while someone on screen demonstrates the opposite.&lt;/p&gt;

&lt;p&gt;What I will remember is finishing. A hackathon project that still makes sense on the Monday after is rare, and this one is small enough to keep. The next step is to run it on a real repository, my own included, and read every comment it writes as if a stranger had written it to me.&lt;/p&gt;

&lt;p&gt;If you maintain a project and want to try it, it is one block in a workflow. Tell me what it says that it should not, or what it misses.&lt;/p&gt;

</description>
      <category>mlhacks</category>
      <category>devchallenge</category>
      <category>hackathon</category>
      <category>opensource</category>
    </item>
    <item>
      <title>Requiem: a crash reporter that writes a eulogy for the program that died</title>
      <dc:creator>James Ngandu</dc:creator>
      <pubDate>Thu, 05 Nov 2026 19:55:10 +0000</pubDate>
      <link>https://dev.to/mukuvi/requiem-a-crash-reporter-that-writes-a-eulogy-for-the-program-that-died-5d6b</link>
      <guid>https://dev.to/mukuvi/requiem-a-crash-reporter-that-writes-a-eulogy-for-the-program-that-died-5d6b</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/mlh-hackathon"&gt;MLH x DEV Writing Challenge&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;Every crash gets a funeral.&lt;/p&gt;

&lt;p&gt;Requiem is a small tool you wrap around any command. When the program dies, Requiem does not just print a traceback and move on. It captures how the process died, works out what killed it, and then writes an obituary in the first person, as if the program were speaking at its own funeral.&lt;/p&gt;

&lt;p&gt;The idea came from a bad habit we all have. Something crashes in production, we copy the stack trace into a chat window, and a few seconds later a cloud model explains the error back to us. Convenient, until you remember what is actually inside a traceback: file paths, hostnames, SQL, auth headers, customer names, and sometimes a token someone left in an environment variable. That is a lot of context to hand to a stranger.&lt;/p&gt;

&lt;p&gt;So Requiem keeps the whole thing on your machine. It runs inference locally with an open-weight model, and before the model sees anything, a coroner pass strips the secrets out.&lt;/p&gt;

&lt;p&gt;It works in four steps:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Capture.&lt;/strong&gt; Run the command under Requiem and record the exit code, the signal, stderr, stdout, and how long it took.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Coroner.&lt;/strong&gt; Parse the traceback, map the signal, read the exact failing line from your source, and redact tokens, keys, emails, home paths and IP addresses.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ghost.&lt;/strong&gt; A local Gemma 3 1B server writes a first-person eulogy and a short diagnosis. If the model is not running, a deterministic obituary is generated instead, so the tool never depends on the network.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Bury.&lt;/strong&gt; The obituary goes into a local JSONL graveyard, and a card is printed to your terminal.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The part I am happiest with is the split of work. Code does what code is good at: parsing, redaction, signal mapping, file reads. The model does the one thing code is bad at, which is voice. Both halves are replaceable. Swap the GGUF for any other model. Change the prompt. Read the 400 lines of standard library Python and know exactly what happened.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;There is no video yet, mostly because the demo is a terminal and the whole thing is more fun to read than to watch. Here is the real output from a demo crash:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;┌─ REQUIEM ──────────────────────────────────────────────────┐
│ requiem-8f3a21  python3 examples/lost_config.py  112ms     │
│ died of KeyError: 'user_id'                                 │
├─────────────────────────────────────────────────────────────┤
│ WHERE                                                       │
│   examples/lost_config.py:14 in load_user                   │
│     return config["users"][user_id]                         │
│ CAUSE                                                       │
│   KeyError: 'user_id' - looked up a missing key in a dict.  │
│ FIX                                                         │
│   Validate input or use .get() with a default.              │
│ EULOGY                                                      │
│   I was only ever asked for keys I had. I never learned to  │
│   say no, and now I say nothing.                            │
│ redacted:0                                                  │
└─────────────────────────────────────────────────────────────┘
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Try it yourself:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;make llama                                     &lt;span class="c"&gt;# start the local model&lt;/span&gt;
&lt;span class="nv"&gt;PYTHONPATH&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="nb"&gt;.&lt;/span&gt; python3 &lt;span class="nt"&gt;-m&lt;/span&gt; requiem run &lt;span class="nt"&gt;--&lt;/span&gt; &amp;lt;&lt;span class="nb"&gt;command&lt;/span&gt;&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;
&lt;span class="nv"&gt;PYTHONPATH&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="nb"&gt;.&lt;/span&gt; python3 &lt;span class="nt"&gt;-m&lt;/span&gt; requiem list
&lt;span class="nv"&gt;PYTHONPATH&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="nb"&gt;.&lt;/span&gt; python3 &lt;span class="nt"&gt;-m&lt;/span&gt; requiem show requiem-xxxxxx
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Add &lt;code&gt;--no-model&lt;/code&gt; and it skips the ghost entirely and prints a deterministic obituary, which is useful on a plane or in CI.&lt;/p&gt;

&lt;p&gt;The repo is &lt;a href="https://github.com/mukuvi/requiem" rel="noopener noreferrer"&gt;github.com/mukuvi/requiem&lt;/a&gt;, MIT licensed. The original write-up for the challenge week is &lt;a href="https://dev.to/mukuvi/requiem-a-crash-post-mortem-that-speaks-as-the-program-that-died-3df1"&gt;here&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Partner Technologies
&lt;/h2&gt;

&lt;p&gt;I did not use a partner API in this project, and that was the point rather than an accident.&lt;/p&gt;

&lt;p&gt;The thing Requiem processes is the most sensitive file a developer owns, and the entire reason it exists is that uploading a crash log to a third party is a bad trade. Adding a cloud call would have made the demo easier and the product pointless.&lt;/p&gt;

&lt;p&gt;What it does use is &lt;strong&gt;Gemma 3 1B&lt;/strong&gt;, an open-weight model, served locally through &lt;code&gt;llama.cpp&lt;/code&gt;. That choice gave me two things a hosted API could not. First, the eulogy never leaves &lt;code&gt;127.0.0.1&lt;/code&gt;. Second, the model is a file. You can swap it for a different GGUF, run it on a weak laptop, or delete it and keep the deterministic path.&lt;/p&gt;

&lt;p&gt;Working with a 1B model also forced better engineering. Small models do not forgive a vague prompt. I had to hand the ghost a structured cause and a clean failing line instead of dumping raw stderr at it, which turned out to be the same discipline that made the redaction pass work. The constraint made the tool simpler, not worse.&lt;/p&gt;

&lt;h2&gt;
  
  
  Hackathon Experience
&lt;/h2&gt;

&lt;p&gt;I built this during &lt;strong&gt;Hacktoberfest 2026&lt;/strong&gt;, the month that MLH and DEV run together, and I checked into &lt;strong&gt;Global Hack Week: Hacktoberfest&lt;/strong&gt; in the middle of it.&lt;/p&gt;

&lt;p&gt;This year was the first one without a pull request target, which sounds like a small rule change and quietly changes the whole month. Without a counter to hit, the only thing left to do is build something you actually want to exist. I ended up making four small tools across October, one for each weekly prompt, all local and all open source. Requiem was the last one, for the Ghost in the Machine week, and it is the one I would keep.&lt;/p&gt;

&lt;p&gt;What I will remember is the rhythm more than a single moment. A new prompt on Monday, a working prototype by the weekend, and a room full of people on the other side of a Discord channel doing the same thing in their own timezone. Building in Nairobi means the livestreams land late, so most of my work happened between eleven at night and two in the morning, with a terminal on one side of the screen and the community feed on the other.&lt;/p&gt;

&lt;p&gt;The other thing I will remember is how much a deadline the size of a week improves a design. Requiem exists in the shape it does because I had six days and no appetite for a framework. No database, no service, no dashboard. A subprocess wrapper, a redactor, a JSONL file and a model. When the month ended I did not have to throw anything away, which is rare for a hackathon project.&lt;/p&gt;

&lt;p&gt;If you have a crash that deserved better than a stack trace, I would love to read its eulogy. The comments are open, and the repo takes issues.&lt;/p&gt;

&lt;p&gt;Thanks to MLH and DEV for the month, and to everyone who read the other three write-ups and left a note.&lt;/p&gt;

</description>
      <category>mlhacks</category>
      <category>devchallenge</category>
      <category>hackathon</category>
      <category>ai</category>
    </item>
    <item>
      <title>One prompt, six boxes, 25/25: building an OTP screen in 30 minutes</title>
      <dc:creator>James Ngandu</dc:creator>
      <pubDate>Thu, 05 Nov 2026 19:53:36 +0000</pubDate>
      <link>https://dev.to/mukuvi/one-prompt-six-boxes-2525-building-an-otp-screen-in-30-minutes-4mco</link>
      <guid>https://dev.to/mukuvi/one-prompt-six-boxes-2525-building-an-otp-screen-in-30-minutes-4mco</guid>
      <description>&lt;p&gt;The challenge was simple to state and hard to finish: build an OTP verification screen with one AI prompt, in 30 minutes, and let a CI pipeline and an AI grader decide how good it is.&lt;/p&gt;

&lt;p&gt;This was hakkahakka26 at droidcon Kenya 2026. The room had 20 teams, a 30-minute clock, and one prompt per team. Teams picked a mobile stack, copied a template, committed the prompt they used as &lt;code&gt;PROMPT.md&lt;/code&gt;, added a screenshot and a demo, and opened a pull request. The repo's Build workflow compiled the app and ran the tests. Then Claude Sonnet graded the submission out of 25 and posted the score on the PR.&lt;/p&gt;

&lt;p&gt;I built the Flutter version. It scored 23/25 on the first pass, then 25/25 after a focused refactor. This is what happened, what the grader cared about, and what I would do differently next time.&lt;/p&gt;

&lt;h2&gt;
  
  
  The screen
&lt;/h2&gt;

&lt;p&gt;The brief asked for a bottom sheet over a dimmed, blurred background. The sheet needed a logo, a title, a subtitle, six 48dp digit boxes, a resend line, and a full-width rounded Submit button.&lt;/p&gt;

&lt;p&gt;The digit boxes had three states:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;filled: dark green circle, white digit&lt;/li&gt;
&lt;li&gt;focused: white circle, dark outline&lt;/li&gt;
&lt;li&gt;empty: light grey circle, dash placeholder&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The interactions were where the real work was:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;the numeric keyboard opens automatically&lt;/li&gt;
&lt;li&gt;typing a digit advances the focus&lt;/li&gt;
&lt;li&gt;backspace on an empty box clears the previous digit&lt;/li&gt;
&lt;li&gt;pasting &lt;code&gt;Your code is 482913&lt;/code&gt; fills &lt;code&gt;482913&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;Submit stays disabled until all six digits are present&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;123456&lt;/code&gt; shows &lt;code&gt;You're in.&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;any other code shows the error, clears the boxes, and focuses the first box&lt;/li&gt;
&lt;li&gt;Resend is disabled for a 30-second countdown, then tappable, and tapping it restarts the countdown&lt;/li&gt;
&lt;li&gt;each box is announced as "Digit N of 6"&lt;/li&gt;
&lt;li&gt;the screen supports dark mode&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The copy had to match the challenge list exactly, even where the reference image differed. That detail matters: the reference said "Enter Your code" and "your cell phone", but the copy list said "Enter your code" and "your phone". The list wins.&lt;/p&gt;

&lt;h2&gt;
  
  
  Six boxes, one text field
&lt;/h2&gt;

&lt;p&gt;The cleanest decision was to use one invisible &lt;code&gt;TextField&lt;/code&gt; as the source of truth and render six boxes from its value. That makes paste and backspace behave like a normal text field, while the UI still looks like six separate inputs.&lt;/p&gt;

&lt;p&gt;The visible state is derived:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight dart"&gt;&lt;code&gt;&lt;span class="kd"&gt;final&lt;/span&gt; &lt;span class="n"&gt;filled&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;index&lt;/span&gt; &lt;span class="p"&gt;&amp;lt;&lt;/span&gt; &lt;span class="n"&gt;digits&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="na"&gt;length&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kd"&gt;final&lt;/span&gt; &lt;span class="n"&gt;focused&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;index&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="n"&gt;digits&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="na"&gt;length&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The hidden field keeps the input honest:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight dart"&gt;&lt;code&gt;&lt;span class="nl"&gt;inputFormatters:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;
  &lt;span class="n"&gt;FilteringTextInputFormatter&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="na"&gt;digitsOnly&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;LengthLimitingTextInputFormatter&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;6&lt;/span&gt;&lt;span class="p"&gt;),&lt;/span&gt;
&lt;span class="p"&gt;],&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The paste case then works without special clipboard code. &lt;code&gt;Your code is 482913&lt;/code&gt; becomes &lt;code&gt;482913&lt;/code&gt;, and the boxes update from the same string. Backspace removes the last digit, which is the same as clearing the previous box and moving the focus back.&lt;/p&gt;

&lt;p&gt;The sheet itself is split into small pieces: &lt;code&gt;OtpScreen&lt;/code&gt; owns state, &lt;code&gt;OtpSheet&lt;/code&gt; owns the layout, &lt;code&gt;OtpBackdrop&lt;/code&gt; owns the dimmed brand background, and &lt;code&gt;OtpInput&lt;/code&gt; owns the six circles. Colours come from &lt;code&gt;assets/colours.json&lt;/code&gt;, so light and dark mode share one source of truth.&lt;/p&gt;

&lt;h2&gt;
  
  
  The pipeline
&lt;/h2&gt;

&lt;p&gt;The submission format is strict. The PR can only touch &lt;code&gt;submissions/&amp;lt;team-name&amp;gt;/&lt;/code&gt;. It needs:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;&lt;code&gt;PROMPT.md&lt;/code&gt;&lt;/li&gt;
&lt;li&gt;&lt;code&gt;TEAM.md&lt;/code&gt;&lt;/li&gt;
&lt;li&gt;&lt;code&gt;screenshot.png&lt;/code&gt;&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;demo.gif&lt;/code&gt; or &lt;code&gt;demo.mp4&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;&lt;code&gt;app/&lt;/code&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The Build workflow detects the stack from &lt;code&gt;pubspec.yaml&lt;/code&gt;, runs &lt;code&gt;flutter build apk --debug&lt;/code&gt;, and runs &lt;code&gt;flutter test&lt;/code&gt;. A separate Grade workflow runs from &lt;code&gt;main&lt;/code&gt;, checks out no PR code, and sends the diff, prompt, screenshot, and build result to Claude Sonnet with a fixed rubric.&lt;/p&gt;

&lt;p&gt;The first score was 23/25:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;acceptance criteria: 10/10&lt;/li&gt;
&lt;li&gt;visual match: 6/6&lt;/li&gt;
&lt;li&gt;code quality: 3/4&lt;/li&gt;
&lt;li&gt;build: 2/2&lt;/li&gt;
&lt;li&gt;tests: 2/2&lt;/li&gt;
&lt;li&gt;prompt length: 0/1&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The feedback was specific. The screen file was doing too much, the fallback palette duplicated the JSON, and the prompt was 862 characters, over the 600-character limit for the point.&lt;/p&gt;

&lt;h2&gt;
  
  
  The second pass
&lt;/h2&gt;

&lt;p&gt;I split the screen into &lt;code&gt;OtpSheet&lt;/code&gt; and &lt;code&gt;OtpBackdrop&lt;/code&gt;, kept the state in &lt;code&gt;_OtpScreenState&lt;/code&gt;, and removed the duplicated fallback palette. The app now loads &lt;code&gt;assets/colours.json&lt;/code&gt; once at startup, and the tests load the same asset. I also tightened the submitted prompt to 600 characters.&lt;/p&gt;

&lt;p&gt;Then I pushed again. The Build workflow stayed green, all 12 tests passed, and the next grade was 25/25.&lt;/p&gt;

&lt;p&gt;The tests cover the parts that are easy to fake:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;the copy and six 48dp boxes&lt;/li&gt;
&lt;li&gt;numeric keyboard autofocus and digit rejection&lt;/li&gt;
&lt;li&gt;auto-advance and backspace&lt;/li&gt;
&lt;li&gt;SMS paste&lt;/li&gt;
&lt;li&gt;Submit gating&lt;/li&gt;
&lt;li&gt;the exact success and error messages&lt;/li&gt;
&lt;li&gt;the 30-second resend countdown and restart&lt;/li&gt;
&lt;li&gt;the &lt;code&gt;Digit N of 6&lt;/code&gt; semantics&lt;/li&gt;
&lt;li&gt;dark mode&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;They are not decoration. The test point is real, and the error path is the easiest place to lose it.&lt;/p&gt;

&lt;h2&gt;
  
  
  The screenshot and demo
&lt;/h2&gt;

&lt;p&gt;I did not have a local Flutter SDK on the machine, so I used GitHub Actions for the whole loop. The repo's Build workflow compiled the APK and ran the tests. A throwaway capture workflow ran a golden test against the widget tree, rendered the empty, partial, filled, success, error, resend, and dark frames, and assembled them into a GIF with ImageMagick.&lt;/p&gt;

&lt;p&gt;That workflow was temporary. The final PR only contains &lt;code&gt;submissions/malhare/&lt;/code&gt;. The screenshot shows the partially filled state with &lt;code&gt;2 1 5&lt;/code&gt;, matching the reference layout. The GIF shows the full interaction.&lt;/p&gt;

&lt;h2&gt;
  
  
  The result
&lt;/h2&gt;

&lt;p&gt;The final submission scored 25/25 with 8/8 checklist items. The PR is here:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/buildthamani/hakka26/pull/24" rel="noopener noreferrer"&gt;https://github.com/buildthamani/hakka26/pull/24&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The repo is here:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/buildthamani/hakka26" rel="noopener noreferrer"&gt;https://github.com/buildthamani/hakka26&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The leaderboard is here:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://buildthamani.github.io/hakka26/" rel="noopener noreferrer"&gt;https://buildthamani.github.io/hakka26/&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The humbling part: 25/25 is the ceiling, and several teams reached it before me. The tie-break is the earliest graded run, so the top-three demo places were already taken by the time my second pass landed. The score is maxed, but the clock is part of the competition.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I learned
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;One prompt can get you a working screen. The last points come from structure, tests, and following the format.&lt;/li&gt;
&lt;li&gt;The prompt length is a scoring constraint. A detailed prompt can be too long for the point.&lt;/li&gt;
&lt;li&gt;The screenshot and demo are part of the submission, not an afterthought.&lt;/li&gt;
&lt;li&gt;A fast CI loop is worth more than a local setup you cannot run.&lt;/li&gt;
&lt;li&gt;Read the rubric before the first push. The grader tells you exactly what it wants.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The code and the PR are public. If you are building an OTP screen, the six-boxes-one-field pattern is a good place to start, and the tests are a good checklist for the edge cases.&lt;/p&gt;

</description>
      <category>flutter</category>
      <category>hackathon</category>
      <category>dart</category>
      <category>mobile</category>
    </item>
    <item>
      <title>Hacktoberfest 2026: four local AI tools and a playground for first-timers</title>
      <dc:creator>James Ngandu</dc:creator>
      <pubDate>Tue, 03 Nov 2026 18:48:26 +0000</pubDate>
      <link>https://dev.to/mukuvi/hacktoberfest-2026-four-local-ai-tools-and-a-playground-for-first-timers-9dj</link>
      <guid>https://dev.to/mukuvi/hacktoberfest-2026-four-local-ai-tools-and-a-playground-for-first-timers-9dj</guid>
      <description>&lt;p&gt;October used to come with a number attached. Four pull requests, six in some years, and a t-shirt at the end of it. This year Hacktoberfest took the number away.&lt;/p&gt;

&lt;p&gt;The organisers were direct about why. Pull requests no longer count toward rewards, because a machine can produce a hundred low-effort ones before lunch, and the maintainers are the ones who have to read them. The month was rebuilt around learning, around events, and around a theme I have not stopped thinking about: AI belongs to everyone.&lt;/p&gt;

&lt;p&gt;I expected to feel robbed. A month without a target sounded like a month without a reason to show up.&lt;/p&gt;

&lt;p&gt;I ended up shipping more than I have in any October before.&lt;/p&gt;

&lt;h2&gt;
  
  
  Open source is where I learned to work
&lt;/h2&gt;

&lt;p&gt;My GitHub bio has said the same thing for years: late nights, early mornings, all for the love of coding. That is not a line I wrote for a profile. It is how the last few years actually went.&lt;/p&gt;

&lt;p&gt;I came up through ALX, writing C by hand before I trusted myself with anything bigger. Pointers, malloc, a shell of my own. Then Python, then shell scripts, then servers that had to stay up while I slept. Almost everything I know came from reading somebody else's repository at midnight and trying to make it run on my own machine.&lt;/p&gt;

&lt;p&gt;That is the part people skip when they describe open source. It is not only about writing code for strangers. It is about being allowed to read how other people think.&lt;/p&gt;

&lt;h2&gt;
  
  
  Four projects, four weeks, everything local
&lt;/h2&gt;

&lt;p&gt;Once the pull request target disappeared, the only useful question left was the one I usually avoid: what do I actually want to build?&lt;/p&gt;

&lt;p&gt;The theme answered it. If AI belongs to everyone, it should run on the laptop of someone who cannot afford an API bill. So every project I built this month runs on the machine in front of me, with open-weight models and no account.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Touch Grass Birder.&lt;/strong&gt; Week one, theme Touch Grass. I hold my phone up for a few seconds and it tells me which bird is singing. BirdNET does the listening through ONNX Runtime, and Gemma 3 1B writes the field note. Audio never leaves the device and there is no key to configure.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/mukuvi/touch-grass-birder" rel="noopener noreferrer"&gt;GitHub&lt;/a&gt; and &lt;a href="https://dev.to/mukuvi/touch-grass-birder-a-bird-call-identifier-that-works-with-no-signal-1hnp"&gt;the write-up&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fg9hb3a22nlhd8inf0q9o.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fg9hb3a22nlhd8inf0q9o.png" alt="Quiet Ledger's ledger view, a local-first receipt and invoice sorter" width="800" height="719"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Quiet Ledger.&lt;/strong&gt; Week two, theme Side Hustle. A receipt sorter for freelancers who bill by the hour and would rather not hand their client data to another SaaS company. Drop in a PDF or a photo, Gemma 3 1B reads it locally, drafts the entry, you confirm it, and it lands in the CSV your accountant asked for.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/mukuvi/quiet-ledger" rel="noopener noreferrer"&gt;GitHub&lt;/a&gt; and &lt;a href="https://dev.to/mukuvi/quiet-ledger-a-receipt-sorter-that-never-uploads-your-receipts-ion"&gt;the write-up&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Dozer.&lt;/strong&gt; A folder tidier that works while I sleep. It reads the Downloads folder one file at a time, files invoices and installers and photos where they belong, renames the files called &lt;code&gt;scan_0043_final.pdf&lt;/code&gt;, and leaves anything it is unsure about exactly where it found it.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/mukuvi/dozer" rel="noopener noreferrer"&gt;GitHub&lt;/a&gt; and &lt;a href="https://dev.to/mukuvi/dozer-a-folder-tidier-that-works-while-you-sleep-6d9"&gt;the write-up&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Requiem.&lt;/strong&gt; Week four, theme Ghost in the Machine. A crash post-mortem that speaks as the program that died. It captures the exit code and the traceback, redacts tokens and paths, then asks a local model to write a first-person eulogy and a short diagnosis. When the model is offline it falls back to a deterministic obituary, which is somehow funnier.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/mukuvi/requiem" rel="noopener noreferrer"&gt;GitHub&lt;/a&gt; and &lt;a href="https://dev.to/mukuvi/requiem-a-crash-post-mortem-that-speaks-as-the-program-that-died-3df1"&gt;the write-up&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Four projects, no cloud bill, no key to leak. A model running on your own laptop is the most literal version of "AI belongs to everyone" that I could build in a month.&lt;/p&gt;

&lt;h2&gt;
  
  
  The badges
&lt;/h2&gt;

&lt;p&gt;I will not pretend the badges do not matter to me.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fholopin.me%2Fjmngandu" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fholopin.me%2Fjmngandu" alt="James Ngandu's Holopin badge wall" width="760" height="239"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;My wall is at &lt;a href="https://www.holopin.io/@jmngandu" rel="noopener noreferrer"&gt;holopin.io/@jmngandu&lt;/a&gt;. Each badge is a small piece of plastic for a profile page, and each one also marks a specific night I can point at. The night ONNX Runtime finally loaded the classifier. The night the receipt parser stopped mistaking totals for dates. The night the eulogy generator made me laugh out loud at one in the morning.&lt;/p&gt;

&lt;p&gt;Three stickers earned a real sticker pack this year, and ten and fifteen opened more. I was not collecting for the pack. I was collecting because a badge is a receipt. It says: I was here, I built the thing, I finished it.&lt;/p&gt;

&lt;h2&gt;
  
  
  The project I did not plan
&lt;/h2&gt;

&lt;p&gt;The best thing I made this month was not on the challenge list.&lt;/p&gt;

&lt;p&gt;People keep asking the same question in the same nervous tone. How do I make my first open source contribution? The honest answer is usually a wall of jargon about forks, upstreams, rebases and maintainers who are too busy to reply.&lt;/p&gt;

&lt;p&gt;So I built a place where none of that can hurt you. &lt;a href="https://free-open.vercel.app/" rel="noopener noreferrer"&gt;FREE&amp;amp;OPEN&lt;/a&gt; is a small React site where your entire first pull request is adding your own name to a JSON file. Fork it, branch it, add your card, open the pull request. Your name appears on the contributors page once it merges.&lt;/p&gt;

&lt;p&gt;The repo is &lt;a href="https://github.com/mukuvi/FREE-OPEN" rel="noopener noreferrer"&gt;github.com/mukuvi/FREE-OPEN&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;It has 17 contributors now, 20 forks, about 300 commits and an MIT licence. Seventeen people who each did the whole loop once: fork, clone, branch, commit, push, pull request, merged. That loop is the door to everything else, and it should not require a stranger to explain it to you first.&lt;/p&gt;

&lt;h2&gt;
  
  
  What this month actually taught me
&lt;/h2&gt;

&lt;p&gt;Smaller is better. Every project above started as one file and one idea, and the ones that survived are the ones I resisted expanding.&lt;/p&gt;

&lt;p&gt;Write the README first. If I cannot explain the tool in three sentences, I do not understand it yet.&lt;/p&gt;

&lt;p&gt;Build for the person you were two years ago. That person is still out there, and they are the best user you will ever have.&lt;/p&gt;

&lt;p&gt;A badge is a receipt, not a scoreboard. The number of them says less than what each one made you learn.&lt;/p&gt;

&lt;p&gt;And the part I keep coming back to is the people. Seventeen names on a page I built. Four projects with write-ups that strangers read. A month of late nights that produced something instead of just consuming.&lt;/p&gt;

&lt;h2&gt;
  
  
  If you are starting
&lt;/h2&gt;

&lt;p&gt;The rules changed this year, but the door is the same. Find a project small enough to understand. Read the CONTRIBUTING file before you touch anything. Open the pull request even though your hands are shaking.&lt;/p&gt;

&lt;p&gt;Or skip the search and start here: &lt;a href="https://github.com/mukuvi/FREE-OPEN" rel="noopener noreferrer"&gt;add your name to FREE&amp;amp;OPEN&lt;/a&gt;. Ten minutes, one file, and forks will stop being scary.&lt;/p&gt;

&lt;p&gt;Everything else is on my GitHub: &lt;a href="https://github.com/mukuvi" rel="noopener noreferrer"&gt;github.com/mukuvi&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Happy Hacktoberfest. A few days late, but the badges just arrived.&lt;/p&gt;

</description>
      <category>hacktoberfest</category>
      <category>opensource</category>
      <category>ai</category>
      <category>beginners</category>
    </item>
    <item>
      <title>Payments were never the hard part: what we are building at FinoQ</title>
      <dc:creator>James Ngandu</dc:creator>
      <pubDate>Tue, 03 Nov 2026 18:22:05 +0000</pubDate>
      <link>https://dev.to/mukuvi/payments-were-never-the-hard-part-what-we-are-building-at-finoq-5375</link>
      <guid>https://dev.to/mukuvi/payments-were-never-the-hard-part-what-we-are-building-at-finoq-5375</guid>
      <description>&lt;p&gt;A shopkeeper in Nairobi closes up on Sunday night. By Monday morning she has three M-Pesa messages for the till, a paybill statement from the bank, one supplier paid in cash, and a notebook with two columns. Her accountant wants statements by Wednesday.&lt;/p&gt;

&lt;p&gt;Nothing on that list is a payment problem. The money moved. It moved fine.&lt;/p&gt;

&lt;p&gt;The problem is that the story of the money now lives in five places, and somebody has to sit down and put it back together before the week starts again.&lt;/p&gt;

&lt;p&gt;That is the work we are building for.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fr4vibbas9khi8i4dtpny.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fr4vibbas9khi8i4dtpny.webp" alt="Two colleagues looking at a laptop together" width="800" height="533"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The part nobody sells you
&lt;/h2&gt;

&lt;p&gt;Payments get all the attention. New wallets, new gateways, new checkout buttons, another QR code on the counter. Every one of them asks a business to move something: move your customers, move your float, move your till number, move your books.&lt;/p&gt;

&lt;p&gt;Most businesses we talk to do not want to move anything. Their paybill works. Their till works. Their bank works. Their accountant has a Zoho Books file that has been tidy for four years. What they want is for all of it to stop being five separate jobs.&lt;/p&gt;

&lt;p&gt;So we built the opposite of another wallet. FinoQ connects the channels a business already uses and puts one readable record underneath them.&lt;/p&gt;

&lt;p&gt;Connect, do not replace. That sentence is the whole product.&lt;/p&gt;

&lt;h2&gt;
  
  
  What FinoQ actually is
&lt;/h2&gt;

&lt;p&gt;Four pieces, one account.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Before the payment.&lt;/strong&gt; We check who you are dealing with. National ID, passport, KRA PIN, company lookups. Then we match a phone number or a bank account to the person who owns it. If the name on the account is not the name you expected, you find out before the money leaves, not after.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Getting paid and paying out.&lt;/strong&gt; Payment links, an M-Pesa prompt on the customer's phone, and the paybill, till or bank you already run. On the way out you get single, bulk and scheduled payouts, with the registered name on the screen first.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;After the payment.&lt;/strong&gt; Collections and payouts land on one double-entry ledger. Sales are fiscalised with KRA eTIMS as they happen. Invoices, payments and expenses mirror into Zoho Books. Revenue, outgoing and net are worked out per channel, not per spreadsheet.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Where you use it.&lt;/strong&gt; The dashboard if you want no code. The API if you want all of it. Same checks, same limits, same events.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fsx52xf8xv1kpwl0d52gp.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fsx52xf8xv1kpwl0d52gp.webp" alt="FinoQ Business home screen showing balances, collections and connected channels" width="800" height="551"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The screenshot we are proudest of
&lt;/h2&gt;

&lt;p&gt;It is not the dashboard. It is the confirm screen.&lt;/p&gt;

&lt;p&gt;You type a phone number, and before you press anything, the app shows you the name registered to that number. Wanjiku Kamau. Not a masked account, not a reference code, not a prayer.&lt;/p&gt;

&lt;p&gt;We built it because sending money to the wrong person is a quiet, expensive, very common kind of pain. The kind you only talk about after it happens. Getting the name in front of someone for two seconds prevents a whole category of support tickets, and a whole category of arguments.&lt;/p&gt;

&lt;p&gt;It also changes how a business pays suppliers. When the name is on screen, an approval means something.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fm10xxbxegaaeof1w1dcz.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fm10xxbxegaaeof1w1dcz.webp" alt="Send money screen with the recipient name and channel check" width="440" height="956"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  One wallet, many countries
&lt;/h2&gt;

&lt;p&gt;On the personal side, the same idea for people rather than businesses. Your wallet is opened with a licensed partner bank through FinoQ, and you use it from one app. Deposit with an M-Pesa prompt, request money with a link that already has the other person's details filled in, then send to mobile money or a bank account across more than thirty African countries.&lt;/p&gt;

&lt;p&gt;Your money is not sitting with us. FinoQ is not a bank and does not hold funds. The accounts are opened and held by licensed partners, and you can see that in the app instead of having to read the terms.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F0mc3968s534d0nf2ujmd.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F0mc3968s534d0nf2ujmd.webp" alt="FinoQ Personal home screen with balance and quick actions" width="455" height="943"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  For the people who will integrate it
&lt;/h2&gt;

&lt;p&gt;The developer story is simple to say and hard to build: one integration, the whole ecosystem.&lt;/p&gt;

&lt;p&gt;One API key. One sandbox. One webhook pipeline with a signing secret you can rotate, and a log that shows every payload and every provider response. A discovery endpoint tells you which channels are available for a country and an amount before you write the request.&lt;/p&gt;

&lt;p&gt;The part we care about most is that you keep your own provider relationships. If you already went through the work of getting M-Pesa credentials, you can attach them. Transactions run under your own merchant account. A payment provider sells you its account. We would rather you kept yours.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl &lt;span class="nt"&gt;-X&lt;/span&gt; POST https://api.finoq.africa/v1/payments &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-u&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$FINOQ_KEY&lt;/span&gt;&lt;span class="s2"&gt;:&lt;/span&gt;&lt;span class="nv"&gt;$FINOQ_SECRET&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-H&lt;/span&gt; &lt;span class="s2"&gt;"Content-Type: application/json"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-d&lt;/span&gt; &lt;span class="s1"&gt;'{
    "amount": 1500,
    "currency": "KES",
    "channel": "mpesa",
    "phone": "+254700000000",
    "reference": "INV-1042",
    "callback_url": "https://yourapp.example/webhooks/finoq"
  }'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;And parity: if a feature exists in the dashboard, it exists on the API. No feature that only a sales call can unlock. That rule costs us discipline every sprint, and it is the reason developers stay.&lt;/p&gt;

&lt;h2&gt;
  
  
  What we are not
&lt;/h2&gt;

&lt;p&gt;We are not your new wallet. Keep the one you have.&lt;/p&gt;

&lt;p&gt;We are not your new accounting software. Keep the books you have, and we will post into them.&lt;/p&gt;

&lt;p&gt;We are not your new eTIMS software. Your sales just get fiscalised as they happen.&lt;/p&gt;

&lt;p&gt;We are not a bank. We do not hold your money, and we are not trying to.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where we are
&lt;/h2&gt;

&lt;p&gt;A small team in Nairobi, building in public enough that you can kick the tyres. The product is live. The API is live. The parts that are still rough are rough in ways we can name, which is more than most.&lt;/p&gt;

&lt;p&gt;If you run a shop, a school, a clinic or a team with three branches and one M-Pesa statement, we would like to hear what your Monday morning actually looks like. If you are building a product on top of payments, we would like to hear what you wish one integration could do.&lt;/p&gt;

&lt;p&gt;The best version of FinoQ is the one where nobody has to think about it.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fyiwfwfc55qo3ecxxxjww.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fyiwfwfc55qo3ecxxxjww.webp" alt="FinoQ Business channel analytics" width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;If any of this is the problem you are sitting with, the door is open at &lt;a href="https://finoq.africa" rel="noopener noreferrer"&gt;finoq.africa&lt;/a&gt;. For the business side, start at &lt;a href="https://business.finoq.africa" rel="noopener noreferrer"&gt;business.finoq.africa&lt;/a&gt;, and the integration guide is at &lt;a href="https://docs.finoq.africa" rel="noopener noreferrer"&gt;docs.finoq.africa&lt;/a&gt;. You can also just message us on WhatsApp at +254 785 077 143. A person answers.&lt;/p&gt;

</description>
      <category>fintech</category>
      <category>payments</category>
      <category>api</category>
      <category>africa</category>
    </item>
    <item>
      <title>The Hidden Engineering Challenges of Building Fintech Across Africa</title>
      <dc:creator>James Ngandu</dc:creator>
      <pubDate>Thu, 29 Oct 2026 14:33:55 +0000</pubDate>
      <link>https://dev.to/mukuvi/the-hidden-engineering-challenges-of-building-fintech-across-africa-5ce0</link>
      <guid>https://dev.to/mukuvi/the-hidden-engineering-challenges-of-building-fintech-across-africa-5ce0</guid>
      <description>&lt;p&gt;Moving money is easy to describe.&lt;/p&gt;

&lt;p&gt;A customer sends money. A payment provider processes it. The recipient gets paid.&lt;/p&gt;

&lt;p&gt;Then you add a second country, a second currency, and a second provider. The description stays the same. The engineering does not.&lt;/p&gt;

&lt;h2&gt;
  
  
  The problem goes beyond payments
&lt;/h2&gt;

&lt;p&gt;Imagine you are building a fintech in Kenya.&lt;/p&gt;

&lt;p&gt;You integrate M-Pesa and a payment provider. Deposits work. Customers can pay merchants and withdraw funds. Everything looks solved.&lt;/p&gt;

&lt;p&gt;Then the business grows.&lt;/p&gt;

&lt;p&gt;You expand into Uganda. You add another provider. Soon you have customers in Nigeria, merchants in three markets, and money moving between currencies and settlement systems that were never designed to talk to each other.&lt;/p&gt;

&lt;p&gt;The questions change:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;How much money is actually available across all providers right now?&lt;/li&gt;
&lt;li&gt;Which transactions settled, and which are still in flight?&lt;/li&gt;
&lt;li&gt;Does our ledger match the balance each provider reports?&lt;/li&gt;
&lt;li&gt;What did foreign exchange cost us this month?&lt;/li&gt;
&lt;li&gt;Can we trace one payment from initiation to final settlement?&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Those sound like finance questions. Most of them are infrastructure questions, and the answers live in your database, not in a spreadsheet.&lt;/p&gt;

&lt;h2&gt;
  
  
  How the money actually moves
&lt;/h2&gt;

&lt;p&gt;A single deposit already passes through more states than most dashboards show.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;CUSTOMER              PROVIDER / MOBILE MONEY        YOUR PLATFORM
   |                            |                          |
   |  sends KES 1,000           |                          |
   |---------------------------&amp;gt;|                          |
   |                            |  webhook: SUCCESS        |
   |                            |-------------------------&amp;gt;|
   |                            |                  ledger entry written
   |                            |                  status: pending
   |                            |                          |
   |                            |  settlement file, T+1    |
   |                            |-------------------------&amp;gt;|
   |                            |                  reconciled against statement
   |                            |                  status: settled
   |                            |                  balance: available
   |                            |                          |
   |                            |  payout request          |
   |                            |&amp;lt;-------------------------|
   |                            |  payout callback         |
   |                            |-------------------------&amp;gt;|
   |                            |                  entry closed
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The customer sees one step. Your system sees five, and each one can fail in a different way.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. Every provider has its own reality
&lt;/h2&gt;

&lt;p&gt;Different providers expose different APIs, status vocabularies, settlement schedules, and reporting formats.&lt;/p&gt;

&lt;p&gt;One reports a transaction as successful before the funds can be withdrawn. Another processes payouts asynchronously. A third sends settlement reports hours later, sometimes in a CSV with a different date format for every market.&lt;/p&gt;

&lt;p&gt;Your system has to absorb those differences without losing track of the underlying transaction.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Provider says&lt;/th&gt;
&lt;th&gt;What it means for your ledger&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;SUCCESS&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;The customer was debited. The money may still be sitting in the provider's float.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;PENDING&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Do not credit the available balance yet. Keep polling or wait for the callback.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;SETTLED&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Funds are in the provider account and can be paid out.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;REVERSED&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Money came back, sometimes days later, sometimes with no webhook at all.&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;This needs more than API clients. It needs an internal representation of financial events that stays true regardless of which provider produced them.&lt;/p&gt;

&lt;p&gt;Practically, that means:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;stable internal references that survive provider ID changes,&lt;/li&gt;
&lt;li&gt;payment status kept separate from settlement status,&lt;/li&gt;
&lt;li&gt;idempotent handling of duplicate webhooks and retries,&lt;/li&gt;
&lt;li&gt;an append only record of everything that changed a balance.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A retry must not create a second payment. A webhook must be verified before it is trusted. A late event must not rewrite a month that finance already closed.&lt;/p&gt;

&lt;p&gt;None of this matters in a demo. All of it matters the first week real money is involved.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Your ledger has to tell the truth
&lt;/h2&gt;

&lt;p&gt;A dashboard can show a balance. That does not make the balance correct.&lt;/p&gt;

&lt;p&gt;Suppose your ledger says KSh 500,000 and the provider reports KSh 485,000.&lt;/p&gt;

&lt;p&gt;Where did the difference go? Pending settlements, transaction fees, refunds, failed payouts, timing, or a genuine break. Without reconciliation, that difference becomes a number somebody has to chase by hand in three dashboards and a spreadsheet.&lt;/p&gt;

&lt;p&gt;A sound ledger keeps immutable entries, stores money in integer minor units instead of floats, and reconciles against provider statements on a schedule instead of when someone notices.&lt;/p&gt;

&lt;p&gt;It also refuses to collapse every balance into one number. We track four:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Available&lt;/strong&gt;: can be paid out right now.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pending&lt;/strong&gt;: collected from the customer but not settled by the provider.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Reserved&lt;/strong&gt;: held for a payout or refund that has not happened yet.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;In transit&lt;/strong&gt;: sent to a provider, confirmation not received.
&lt;/li&gt;
&lt;/ul&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="kd"&gt;type&lt;/span&gt; &lt;span class="nx"&gt;Balance&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="na"&gt;currency&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kr"&gt;string&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="nl"&gt;available&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kr"&gt;number&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="c1"&gt;// minor units, never floats&lt;/span&gt;
  &lt;span class="nl"&gt;pending&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kr"&gt;number&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="nl"&gt;reserved&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kr"&gt;number&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="nl"&gt;inTransit&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kr"&gt;number&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;};&lt;/span&gt;

&lt;span class="c1"&gt;// The dashboard shows one number. This is the explanation behind it.&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;kes&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nx"&gt;Balance&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="na"&gt;currency&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;KES&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="na"&gt;available&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;412&lt;/span&gt;&lt;span class="nx"&gt;_000_00&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="na"&gt;pending&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;58&lt;/span&gt;&lt;span class="nx"&gt;_000_00&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="na"&gt;reserved&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;15&lt;/span&gt;&lt;span class="nx"&gt;_000_00&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="na"&gt;inTransit&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;15&lt;/span&gt;&lt;span class="nx"&gt;_000_00&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;};&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Add those four and you get the KSh 500,000. Take the pending and in transit money out and you get the KSh 470,000 that is genuinely spendable. Neither number is the truth on its own. The breakdown is.&lt;/p&gt;

&lt;p&gt;Reconciliation is what keeps that breakdown honest:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;  provider A  --&amp;gt;  webhooks        --+
  provider B  --&amp;gt;  status polling  --+--&amp;gt;  transaction state  --+
  provider C  --&amp;gt;  settlement file --+                          |
                                                               v
                                                    +---------------------+
                                                    |   internal ledger   |
                                                    +---------------------+
                                                               |
  provider statements  --------------------------------------&amp;gt; |
                                                               v
                                                    +---------------------+
                                                    |    match engine     |
                                                    +---------------------+
                                                       |      |         |
                                                    matched  timing    break
                                                    close    wait    investigate
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Three outcomes, and only one of them is a bug. Timing differences resolve themselves. Real breaks need a person, an audit trail, and enough context to answer "what happened to this money" without opening five tabs.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. Cross border adds a second set of numbers
&lt;/h2&gt;

&lt;p&gt;Within one currency, a payment has one amount. Across borders, it has several.&lt;/p&gt;

&lt;p&gt;A customer pays in Kenyan shillings and the recipient is paid in naira. Between those two events sit an exchange rate, a spread, provider fees, and time. The rate you quoted is not always the rate you got.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;KES 10,000 collected
     |
     | collection fee 1.5%                  -150 KES
     v
KES 9,850 converted
     |
     | quoted rate 11.90                    -&amp;gt; 117,215 NGN
     | executed rate 11.72 (rate moved)     -&amp;gt; 115,442 NGN
     v
NGN 115,442
     |
     | payout fee                           -1,200 NGN
     v
NGN 114,242 delivered
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The transaction succeeded. The recipient got paid. And the business still lost money it never budgeted for, because the spread moved while the payment was in flight.&lt;/p&gt;

&lt;p&gt;That is why the system has to track the quoted rate, the executed rate, the fees actually charged, and the final settlement amount as separate facts. Merge them into one "amount sent" field and you can no longer explain your own margin.&lt;/p&gt;

&lt;p&gt;At scale, exposure matters as much as volume. A business moving money across four currencies needs to know what it is holding, what it is exposed to, and what each corridor actually costs.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. Observability means watching the money, not just the service
&lt;/h2&gt;

&lt;p&gt;In a normal application, a healthy dashboard means the API is up and latency is fine.&lt;/p&gt;

&lt;p&gt;In financial infrastructure, that is the least interesting signal. You also need to know whether money is moving correctly.&lt;/p&gt;

&lt;p&gt;The signals that earn their place on the wall:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Unmatched transactions and the size of each reconciliation break.&lt;/li&gt;
&lt;li&gt;Payment success and failure rates, split by provider and corridor.&lt;/li&gt;
&lt;li&gt;Settlement delays and payouts that have not confirmed.&lt;/li&gt;
&lt;li&gt;Fees and conversion costs by currency.&lt;/li&gt;
&lt;li&gt;Sudden changes in available liquidity.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;These are the numbers that turn a vague "something feels off" into a specific investigation.&lt;/p&gt;

&lt;p&gt;A payment API returning HTTP 200 means the request worked. It says nothing about whether the money arrived.&lt;/p&gt;

&lt;h2&gt;
  
  
  5. The real work is connecting it
&lt;/h2&gt;

&lt;p&gt;Payment integrations, ledgers, reconciliation, FX records, and treasury reporting are usually built as separate systems by separate teams.&lt;/p&gt;

&lt;p&gt;They describe the same financial operation from different angles, and the value appears when they agree.&lt;/p&gt;

&lt;p&gt;That means building for questions, not just screens:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Can every balance be traced back to the entries that produced it?&lt;/li&gt;
&lt;li&gt;Can every entry be traced back to a provider event?&lt;/li&gt;
&lt;li&gt;Can a break be investigated without leaving the system?&lt;/li&gt;
&lt;li&gt;Can an operator answer "where is this money" in one place?&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This matters most for early stage fintechs. They need dependable financial operations long before they can afford a treasury team to assemble the picture by hand.&lt;/p&gt;

&lt;h2&gt;
  
  
  What we are building at FinoQ
&lt;/h2&gt;

&lt;p&gt;At &lt;a href="https://finoq.africa/" rel="noopener noreferrer"&gt;FinoQ&lt;/a&gt;, we are building financial infrastructure that gives fintech businesses visibility into their money as they scale across Africa.&lt;/p&gt;

&lt;p&gt;The goal is to connect provider liquidity, internal records, reconciliation, FX exposure, and fees so a business can see its financial position in one place, and explain it.&lt;/p&gt;

&lt;p&gt;Because moving money is only one part of a reliable financial product. You also need to know where the money is, what happened to it, what it cost, and whether your records match reality.&lt;/p&gt;

&lt;h2&gt;
  
  
  Final thoughts
&lt;/h2&gt;

&lt;p&gt;Africa's fintech opportunity is not only about faster payments or more customers. It is also about the infrastructure that keeps increasingly complex operations honest as companies expand across borders.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Better rails got us here. Better financial visibility is what carries the next stage.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;That is an engineering problem, and a good one to work on.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;This post started as a conversation on &lt;a href="https://www.linkedin.com/feed/update/urn:li:activity:7521480031303380992/" rel="noopener noreferrer"&gt;LinkedIn&lt;/a&gt;. FinoQ also writes about &lt;a href="https://www.linkedin.com/posts/finoq-africa_fintech-treasury-africa-activity-7521214417338810368-dGCS" rel="noopener noreferrer"&gt;fintech treasury in Africa&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  Further reading
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://dev.to/weston_carnes_d580b505e0c/designing-a-cross-border-payment-system-2bi2"&gt;Designing a cross-border payment system&lt;/a&gt; by Weston Carnes&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://dev.to/0xenx/why-your-order-system-and-your-payment-provider-disagree-1ob"&gt;Why your order system and your payment provider disagree&lt;/a&gt; by 0xenx&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://dev.to/rahmannugar/reconcilation-in-payment-systems-4b64"&gt;Reconciliation in payment systems&lt;/a&gt; by Rahman Nugar&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://dev.to/adodanieln/we-built-a-stripe-for-african-mobile-money-then-discovered-why-it-couldnt-exist-3ica"&gt;We built a Stripe for African mobile money, then discovered why it could not exist&lt;/a&gt; by adodanieln&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>fintech</category>
      <category>payments</category>
      <category>architecture</category>
      <category>africa</category>
    </item>
    <item>
      <title>Requiem: a crash post-mortem that speaks as the program that died</title>
      <dc:creator>James Ngandu</dc:creator>
      <pubDate>Tue, 27 Oct 2026 08:04:36 +0000</pubDate>
      <link>https://dev.to/mukuvi/requiem-a-crash-post-mortem-that-speaks-as-the-program-that-died-3df1</link>
      <guid>https://dev.to/mukuvi/requiem-a-crash-post-mortem-that-speaks-as-the-program-that-died-3df1</guid>
      <description>&lt;p&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week4-2026-10-26"&gt;Hacktoberfest Open Source AI Challenge: Week 4 - Ghost in the Machine&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I built
&lt;/h2&gt;

&lt;p&gt;Requiem is a CLI that gives every crashed program a funeral. When a command fails, Requiem produces an obituary card with the cause of death, a suggested fix, and a first person eulogy written as if the program itself were speaking.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/mukuvi/requiem" rel="noopener noreferrer"&gt;GitHub repo&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nv"&gt;PYTHONPATH&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="nb"&gt;.&lt;/span&gt; python3 &lt;span class="nt"&gt;-m&lt;/span&gt; requiem run &lt;span class="nt"&gt;--&lt;/span&gt; python3 examples/lost_config.py
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;┌─ REQUIEM ─────────────────────────────────────────────────────┐
│ requiem-733bae  python3 examples/lost_config.py  132ms       │
│ died of missing key 'bob' in config dictionary.              │
├──────────────────────────────────────────────────────────────┤
│ WHERE                                                        │
│   examples/lost_config.py:14 in load_user                   │
│     user = config["users"][user_id]                          │
│ CAUSE                                                        │
│   missing key 'bob' in config dictionary.                    │
│ FIX                                                          │
│   add 'bob' to config dictionary.                            │
│ EULOGY                                                       │
│   I was a program. I existed to process requests, to         │
│   organize data.                                             │
└──────────────────────────────────────────────────────────────┘
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A segfault gets the same treatment:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;make examples/segv &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="nv"&gt;PYTHONPATH&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="nb"&gt;.&lt;/span&gt; python3 &lt;span class="nt"&gt;-m&lt;/span&gt; requiem run &lt;span class="nt"&gt;--&lt;/span&gt; ./examples/segv
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;┌─ REQUIEM ─────────────────────────────────────────────────────┐
│ requiem-e9b362  ./examples/segv  1ms                         │
│ died of segmentation fault                                   │
├──────────────────────────────────────────────────────────────┤
│ WHERE                                                        │
│   killed by SIGSEGV                                          │
│ CAUSE                                                        │
│   segmentation fault                                         │
│ FIX                                                          │
│   null                                                       │
│ EULOGY                                                       │
│   I trusted a pointer I had no right to trust. Now I return  │
│   to /dev/null.                                              │
└──────────────────────────────────────────────────────────────┘
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  How it works
&lt;/h2&gt;

&lt;p&gt;Requiem is a hybrid: code does what code must, the model gives the program a voice.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Capture runs the command with subprocess and records exit code, signal, duration, and the last chunk of stderr and stdout.&lt;/li&gt;
&lt;li&gt;Coroner parses Python tracebacks, maps signals to names, reads the failing source line off disk, and redacts secrets before anything leaves the process.&lt;/li&gt;
&lt;li&gt;Ghost talks to a local llama-server running Gemma 3 1B via llama.cpp. It asks for a terse diagnosis and a first person eulogy, caches results, and falls back to deterministic output if the model is offline.&lt;/li&gt;
&lt;li&gt;Graveyard stores every obituary in &lt;code&gt;data/obituaries.jsonl&lt;/code&gt;. Use &lt;code&gt;requiem list&lt;/code&gt; and &lt;code&gt;requiem show &amp;lt;id&amp;gt;&lt;/code&gt; to browse them. The terminal renders a clean obituary card.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;All of this is stdlib Python, about 400 lines. No API keys. No network calls when using &lt;code&gt;--no-model&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why open innovation matters
&lt;/h2&gt;

&lt;p&gt;Crash logs are sensitive. A traceback can contain file paths, hostnames, SQL, auth headers, customer data, or tokens. Sending that to a cloud crash explainer leaks private information. Requiem runs inference on your own machine and redacts secrets before the model ever sees them.&lt;/p&gt;

&lt;p&gt;It works offline. The model, prompt, logic, and examples are all local. Swap the GGUF, edit the prompt, read every line. With &lt;code&gt;--no-model&lt;/code&gt; it is fully deterministic and works in air gapped environments.&lt;/p&gt;

&lt;p&gt;It is inspectable. Redaction rules, traceback parsing, and the eulogy prompt are all visible and easy to change. Open tools let you reason about your failures without handing them to a third party.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize categories
&lt;/h2&gt;

&lt;p&gt;Best Use of Gemma - Gemma 3 1B runs locally through llama.cpp and llama-server to write the eulogy and diagnosis.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/mukuvi/requiem
&lt;span class="nb"&gt;cd &lt;/span&gt;requiem
make demo
make &lt;span class="nb"&gt;test&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Every crash gets a funeral.&lt;/p&gt;

</description>
      <category>hf26challenge</category>
      <category>devchallenge</category>
      <category>ai</category>
      <category>opensource</category>
    </item>
    <item>
      <title>Dozer: a folder tidier that works while you sleep</title>
      <dc:creator>James Ngandu</dc:creator>
      <pubDate>Thu, 22 Oct 2026 11:29:26 +0000</pubDate>
      <link>https://dev.to/mukuvi/dozer-a-folder-tidier-that-works-while-you-sleep-6d9</link>
      <guid>https://dev.to/mukuvi/dozer-a-folder-tidier-that-works-while-you-sleep-6d9</guid>
      <description>&lt;p&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19"&gt;Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Frekd9qtu4ql19nzc0gjo.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Frekd9qtu4ql19nzc0gjo.png" alt="Dozer's plan view: 13 files to file, 6 left alone, 4 needed the model" width="800" height="975"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;My Downloads folder is not a storage problem. It is a stack of small decisions I keep postponing. Every file asks a small question: is this an invoice, a screenshot, something I should keep at all? Answering forty of them is not hard. It is just relentless.&lt;/p&gt;

&lt;p&gt;Dozer reads a folder one level deep, one file at a time, with Gemma 3 1B running on the same machine through &lt;code&gt;llama.cpp&lt;/code&gt;. Then it:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Files each file into &lt;code&gt;Invoices/&lt;/code&gt;, &lt;code&gt;Documents/&lt;/code&gt;, &lt;code&gt;Reading/&lt;/code&gt;, &lt;code&gt;Photos/&lt;/code&gt;, &lt;code&gt;Screenshots/&lt;/code&gt;, &lt;code&gt;Installers/&lt;/code&gt;, &lt;code&gt;Archives/&lt;/code&gt;, &lt;code&gt;Code/&lt;/code&gt; or &lt;code&gt;Duplicates/&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Renames only the files whose names carry no information. &lt;code&gt;a1b2c3d4e5f6a7b8.pdf&lt;/code&gt; became &lt;code&gt;2026-10-08_certificate-of-participation.pdf&lt;/code&gt;. &lt;code&gt;kasha-hardware-invoice.pdf&lt;/code&gt; was left exactly as it was.&lt;/li&gt;
&lt;li&gt;Leaves a file alone when it is not sure, when it is empty, when the type is unknown, or when you touched it in the last ten minutes.&lt;/li&gt;
&lt;li&gt;Never deletes and never overwrites. A name collision gets a &lt;code&gt;-2&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Writes every move to &lt;code&gt;data/runs.jsonl&lt;/code&gt;, and &lt;code&gt;undo&lt;/code&gt; puts the whole run back.&lt;/li&gt;
&lt;li&gt;Works at a pace you choose, with a budget, so an interrupted night resumes where it stopped.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;It is for anyone whose Downloads folder has become the place where decisions go to wait. You set it going before bed, and the only thing waiting for you in the morning is a short note.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;The whole thing is three commands:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;make &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; make models &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; make llama
make demo     &lt;span class="c"&gt;# dashboard on a throwaway copy&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;Or against your real folder:&lt;br&gt;
&lt;/p&gt;
&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;make scan &lt;span class="nv"&gt;ROOT&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;~/Downloads
make night &lt;span class="nv"&gt;ROOT&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;~/Downloads
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;This is a real scan of the 19-file folder that ships in &lt;code&gt;samples/&lt;/code&gt;, trimmed to keep it readable:&lt;br&gt;
&lt;/p&gt;
&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Dozer · /tmp/dozer-demo
19 files · 13 to file · 6 left alone · 9 folders
asked the local model about 4 of 19 files, 4 answered from cache

DOCUMENTS/
  2026-10-08_certificate-of-participation.pdf
    from a1b2c3d4e5f6a7b8.pdf · 95% · The text is a certificate, which is a document.
  2026-09-22_tenancy-agreement.pdf
    from download (3).pdf · 80% · the text reads like a tenancy agreement
  random_notes.txt
    80% · the text reads like meeting notes

INVOICES/
  2026-10-12_tax-invoice.pdf
    from scan_0042.pdf · 85% · the text reads like a tax invoice
  statement-oct.pdf
    85% · the text reads like an account statement

READING/
  doc_8831.pdf
    80% · the text reads like a book

LEFT ALONE
  data_dump.csv - I'm not sure what this is.
  empty.txt - empty file
  notes.txt - This file contains a list of vegetables and spices. It's not a standard document type.
  recent-scan.pdf - touched 1 minute ago, still in use
  ubuntu-notes.md - The text is too short to determine the category.
  weird.xyz - not a type Dozer knows
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;Then the overnight pass:&lt;br&gt;
&lt;/p&gt;
&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;$ make night ROOT=/tmp/dozer-demo PACE=2
[1/13] a1b2c3d4e5f6a7b8.pdf -&amp;gt; Documents/2026-10-08_certificate-of-participation.pdf
[2/13] doc_8831.pdf -&amp;gt; Reading/doc_8831.pdf
[3/13] download (3).pdf -&amp;gt; Documents/2026-09-22_tenancy-agreement.pdf
...
[13/13] VSCodeUserSetup-x64-1.94.2.exe -&amp;gt; Installers/VSCodeUserSetup-x64-1.94.2.exe

MORNING NOTE
  I left 6 files alone. I filed 13 files into 3 Documents, 2 Invoices, 2
  Reading. 4 of them needed the model and the rest were rules. The pass took 24
  seconds and nothing left this machine.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;And the folder afterwards:&lt;br&gt;
&lt;/p&gt;
&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;/tmp/dozer-demo
├── Archives/        node-v22.11.0-linux-x64.tar.xz
├── Code/            scrape.py
├── Documents/       2026-09-22_tenancy-agreement.pdf
│                    2026-10-08_certificate-of-participation.pdf
│                    random_notes.txt
├── Duplicates/      IMG_20260918_0712 (1).heic
├── Installers/      VSCodeUserSetup-x64-1.94.2.exe
├── Invoices/        2026-10-12_tax-invoice.pdf
│                    statement-oct.pdf
├── Photos/          IMG_20260918_0712.heic
├── Reading/         doc_8831.pdf
│                    thinking-fast-and-slow.epub
├── Screenshots/     Screenshot 2026-10-20 at 09.14.02.png
├── data_dump.csv
├── empty.txt
├── notes.txt
├── recent-scan.pdf
├── ubuntu-notes.md
└── weird.xyz
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;The six files still in the root are the point.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F1kwnlh3hn5n8iboeo2ld.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F1kwnlh3hn5n8iboeo2ld.png" alt="The morning note" width="799" height="612"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;


&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/mukuvi" rel="noopener noreferrer"&gt;
        mukuvi
      &lt;/a&gt; / &lt;a href="https://github.com/mukuvi/dozer" rel="noopener noreferrer"&gt;
        dozer
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      A folder tidier that works while you sleep, with Gemma 3 1B running locally.
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;Dozer&lt;/h1&gt;
&lt;/div&gt;

&lt;p&gt;A folder tidier that works while you sleep, with an open-weight model at the core.&lt;/p&gt;

&lt;p&gt;Dozer reads your Downloads one file at a time with &lt;strong&gt;Gemma 3 1B&lt;/strong&gt; running through
&lt;code&gt;llama.cpp&lt;/code&gt; on your own machine. It files invoices, documents, installers, photos
duplicates and the rest into folders, renames the files that have useless names, and
leaves everything it is not sure about exactly where it is.&lt;/p&gt;
&lt;p&gt;No API key. No upload. No account.&lt;/p&gt;
&lt;p&gt;Write-up: &lt;a href="https://dev.to/mukuvi/dozer-a-folder-tidier-that-works-while-you-sleep-6d9" rel="nofollow"&gt;Dozer: a folder tidier that works while you sleep&lt;/a&gt;.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;What it actually does&lt;/h2&gt;
&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;Files the mess in one folder, one level deep.&lt;/li&gt;
&lt;li&gt;Renames only files whose names carry no information (&lt;code&gt;scan_0042.pdf&lt;/code&gt;, &lt;code&gt;IMG_2301.HEIC&lt;/code&gt;
&lt;code&gt;download (3).pdf&lt;/code&gt;). A name a human chose is left alone.&lt;/li&gt;
&lt;li&gt;Keeps identical files in &lt;code&gt;Duplicates/&lt;/code&gt; instead of deleting anything.&lt;/li&gt;
&lt;li&gt;Leaves files alone when it is not sure, when the file is empty, when the type is
unknown, or…&lt;/li&gt;
&lt;/ul&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/mukuvi/dozer" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;MIT licensed, 27 tests that run without a model. &lt;code&gt;make test&lt;/code&gt; on a clean checkout.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;



&lt;pre data-lang="mermaid"&gt;&lt;code&gt;flowchart LR
  A[Folder, one level deep] --&amp;gt; B[hash + short text preview]
  B --&amp;gt; C{extension or text rule?}
  C --&amp;gt;|yes| D[category + confidence]
  C --&amp;gt;|no| E[Gemma 3 1B in a warm llama-server]
  E --&amp;gt; D
  D --&amp;gt; F{confidence &amp;gt;= 0.75?}
  F --&amp;gt;|yes| G[folder + safe name]
  F --&amp;gt;|no| H[left alone, with the reason]
  G --&amp;gt; I[apply one file at a time]
  I --&amp;gt; J[(data/runs.jsonl)]
  J --&amp;gt; K[undo]&lt;/code&gt;&lt;/pre&gt;



&lt;p&gt;Three decisions make a 1B model trustworthy enough to move files.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Rules first.&lt;/strong&gt; An &lt;code&gt;.exe&lt;/code&gt; is an installer. A document with &lt;code&gt;Tax Invoice&lt;/code&gt; and &lt;code&gt;Amount due&lt;/code&gt; in the text is an invoice. In the scan above, 15 of 19 files never reached the model. Rules are free, instant and predictable, and they leave the model only the cases that are actually ambiguous.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;A confidence floor.&lt;/strong&gt; Below 0.75 Dozer does nothing and says why. &lt;code&gt;data_dump.csv&lt;/code&gt; is still in the folder because the model answered &lt;code&gt;other&lt;/code&gt; at 0.0 confidence and said "I'm not sure what this is." That is a correct answer. A tidier that guesses is worse than a folder.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;A warm server, not a subprocess per file.&lt;/strong&gt; Every call goes to one &lt;code&gt;llama-server&lt;/code&gt; on localhost, with the prompt prefix cached. On this 2015 dual-core laptop the difference is large. Loading the model per file costs more than the answer, and the shared prefix means the instructions are evaluated once per run instead of once per file. A fresh ambiguous file takes about thirty seconds here. The first one in a run also pays for loading 806 MB.&lt;/p&gt;

&lt;p&gt;Model answers are also cached by content hash, so a second pass over a folder you already scanned costs nothing. That is why the demo above says "4 needed the model, 4 answered from cache."&lt;/p&gt;

&lt;p&gt;The safety rules are simple. No symlinks, no subfolders, no deletion, no overwrite, and no rename of a name a person wrote. &lt;code&gt;undo&lt;/code&gt; reverses the last run in order and refuses to overwrite a file you put back yourself.&lt;/p&gt;

&lt;h3&gt;
  
  
  What it got wrong
&lt;/h3&gt;

&lt;p&gt;Three things, and they are the interesting part.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It called a book chapter a document.&lt;/strong&gt; &lt;code&gt;doc_8831.pdf&lt;/code&gt; starts with "Chapter 3: Trees and Recursion", and the model filed it under Documents at 95% confidence. The fix was one rule, not a bigger model or a longer prompt: text with &lt;code&gt;Chapter N&lt;/code&gt;, &lt;code&gt;Table of Contents&lt;/code&gt; or an ISBN is reading. Rules are how you correct a small model.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The &lt;code&gt;kind&lt;/code&gt; field was noisy.&lt;/strong&gt; Asked about a certificate, it returned &lt;code&gt;kind: "tax invoice"&lt;/code&gt;. The prompt's examples were doing more work than the instruction, so I removed them.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It invented file names in the morning note.&lt;/strong&gt; I wanted the model to write the note you see at 7am. It produced "I left 6 files, including the one named 'Project Alpha'". There is no Project Alpha in the folder. The note is a template now.&lt;/p&gt;

&lt;p&gt;That is the rule I ended up with. The model answers one narrow question, says how sure it is, and code checks the answer. It does not get to tell me what happened.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;The folder never leaves the machine.&lt;/strong&gt; Downloads is where bank statements, contracts, ID scans and photos of your kitchen end up. A closed sorter has to upload them to read them. Running the weights locally means there is no trust boundary to audit, because there is no boundary.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It works with no network.&lt;/strong&gt; After the one-time model download, the scanner, the rules, the model and the log are all on disk. On a metered or unreliable connection that matters more than any feature.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;You can change the sorter.&lt;/strong&gt; Swapping Gemma for a bigger model, a different quantisation, or a model fine-tuned on your own vendor names is a path and an environment variable. The prompt is a string in the repo. When it behaved badly I could read it and fix it the same afternoon.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It costs nothing per file.&lt;/strong&gt; No per-document price, no rate limit, no key that expires. The run above cost nothing, which matters most for the person with a full folder and no budget for another subscription.&lt;/p&gt;

&lt;h2&gt;
  
  
  My Agent Session
&lt;/h2&gt;

&lt;p&gt;I built this with an agent, and the session is saved. It includes the benchmark that pushed me from a subprocess per file to a warm server, and the moment the model invented a file name in the morning note:&lt;/p&gt;


&lt;div class="ltag-agent-session"&gt;
  &lt;div class="agent-session-header"&gt;
    
    &lt;span class="agent-session-tool-icon-badge" title="OpenCode"&gt;
&lt;/span&gt;
    &lt;span class="agent-session-title"&gt;Building Dozer: a local folder tidier for Hacktoberfest Week 3&lt;/span&gt;
  &lt;/div&gt;

  &lt;div class="agent-session-scroll"&gt;

      &lt;div class="agent-session-message agent-session-user"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-user"&gt;
          You
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Proceed with the Hacktoberfest Week 3 challenge (Slow and Steady). Build something with open-source AI that rewards laziness.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Plan: a local folder tidier called Dozer. Gemma 3 1B through llama.cpp decides where a file belongs; deterministic rules handle the obvious cases; anything below 0.75 confidence is left exactly where it is. Python + FastAPI, JSONL log, undo, and a paced overnight mode.&lt;/p&gt;

                &lt;/div&gt;
              &lt;div class="agent-session-tool-call"&gt;
                
                  &lt;span class="agent-session-chevron"&gt;▸&lt;/span&gt;
                  &lt;span class="tool-name tool-name-getchallenges"&gt;get_challenges&lt;/span&gt;
                    &lt;span class="tool-input-preview"&gt;Week 3 details&lt;/span&gt;
                
                &lt;div class="agent-session-tool-detail"&gt;
                    &lt;div class="tool-section"&gt;
                      &lt;div class="tool-section-label"&gt;Input&lt;/div&gt;
                      &lt;pre class="agent-session-pre"&gt;&lt;code&gt;Week 3 details&lt;/code&gt;&lt;/pre&gt;
                    &lt;/div&gt;
                    &lt;div class="tool-section"&gt;
                      &lt;div class="tool-section-label"&gt;Output&lt;/div&gt;
                      &lt;pre class="agent-session-pre"&gt;&lt;code&gt;Week 3 runs Oct 19-25, theme 'Slow and Steady': build something that rewards laziness. Judging: writing quality weighted most, then theme relevance, creativity, technical execution.&lt;/code&gt;&lt;/pre&gt;
                    &lt;/div&gt;
                &lt;/div&gt;
              &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;First I checked the hardware. An i5-5200U with 4 threads, and the machine was already loaded. A quick llama-bench run showed 3.81 t/s prompt and 0.19 t/s generation. Loading the 806 MB model per file with llama-cli would make the tool unusable, so the model layer became a warm llama-server with prefix caching instead.&lt;/p&gt;

                &lt;/div&gt;
              &lt;div class="agent-session-tool-call"&gt;
                
                  &lt;span class="agent-session-chevron"&gt;▸&lt;/span&gt;
                  &lt;span class="tool-name tool-name-llamabench"&gt;llama-bench&lt;/span&gt;
                    &lt;span class="tool-input-preview"&gt;gemma-3-1b-it-Q4_K_M, 4 threads, pp32/tg8&lt;/span&gt;
                
                &lt;div class="agent-session-tool-detail"&gt;
                    &lt;div class="tool-section"&gt;
                      &lt;div class="tool-section-label"&gt;Input&lt;/div&gt;
                      &lt;pre class="agent-session-pre"&gt;&lt;code&gt;gemma-3-1b-it-Q4_K_M, 4 threads, pp32/tg8&lt;/code&gt;&lt;/pre&gt;
                    &lt;/div&gt;
                    &lt;div class="tool-section"&gt;
                      &lt;div class="tool-section-label"&gt;Output&lt;/div&gt;
                      &lt;pre class="agent-session-pre"&gt;&lt;code&gt;pp32: 3.81 t/s, tg8: 0.19 t/s&lt;/code&gt;&lt;/pre&gt;
                    &lt;/div&gt;
                &lt;/div&gt;
              &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;The core is plan.py. Extension and text rules run first, so most files never reach the model. Rules also produce the safe filenames: only names with no information (scan_0042, IMG_2301, a 16-character hash) get renamed, and only when the model or a rule has a title to offer. Duplicates are compared by SHA-256 and moved to Duplicates/ instead of being deleted.&lt;/p&gt;

                &lt;/div&gt;
              &lt;div class="agent-session-tool-call"&gt;
                
                  &lt;span class="agent-session-chevron"&gt;▸&lt;/span&gt;
                  &lt;span class="tool-name tool-name-pytest"&gt;pytest&lt;/span&gt;
                    &lt;span class="tool-input-preview"&gt;tests/test_plan.py&lt;/span&gt;
                
                &lt;div class="agent-session-tool-detail"&gt;
                    &lt;div class="tool-section"&gt;
                      &lt;div class="tool-section-label"&gt;Input&lt;/div&gt;
                      &lt;pre class="agent-session-pre"&gt;&lt;code&gt;tests/test_plan.py&lt;/code&gt;&lt;/pre&gt;
                    &lt;/div&gt;
                    &lt;div class="tool-section"&gt;
                      &lt;div class="tool-section-label"&gt;Output&lt;/div&gt;
                      &lt;pre class="agent-session-pre"&gt;&lt;code&gt;Failed: two tests assumed a .pdf extension meant pdftotext could read plain text. Fixed by using the real sample PDFs, and an autouse fixture that keeps tests away from the local model.&lt;/code&gt;&lt;/pre&gt;
                    &lt;/div&gt;
                &lt;/div&gt;
              &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Then the first real scan. The model called a book chapter (doc_8831.pdf, 'Chapter 3: Trees and Recursion') a Document at 95% confidence, and its 'why' text was boilerplate. I did not reach for a bigger model. I added one rule: text with 'Chapter N', 'Table of Contents' or an ISBN is Reading. That is the lesson the post is built around.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;I also tried letting the model write the morning note. It produced: "I left 6 files, including the one named 'Project Alpha'" - there is no Project Alpha in the folder. It invented file names. The note became a deterministic template, and the model is now trusted with exactly one decision: the category.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Final run on the 19-file sample folder: 13 filed, 6 left alone, 9 folders, 3 renames, 4 files needed the model. Paced at 2 seconds per file, the pass took 24 seconds. A fresh ambiguous file costs about 28 seconds on this laptop, and answers are cached by content hash so the next scan is instant.&lt;/p&gt;

                &lt;/div&gt;
              &lt;div class="agent-session-tool-call"&gt;
                
                  &lt;span class="agent-session-chevron"&gt;▸&lt;/span&gt;
                  &lt;span class="tool-name tool-name-clipynight"&gt;cli.py night&lt;/span&gt;
                    &lt;span class="tool-input-preview"&gt;night /tmp/dozer-demo --pace 2&lt;/span&gt;
                
                &lt;div class="agent-session-tool-detail"&gt;
                    &lt;div class="tool-section"&gt;
                      &lt;div class="tool-section-label"&gt;Input&lt;/div&gt;
                      &lt;pre class="agent-session-pre"&gt;&lt;code&gt;night /tmp/dozer-demo --pace 2&lt;/code&gt;&lt;/pre&gt;
                    &lt;/div&gt;
                    &lt;div class="tool-section"&gt;
                      &lt;div class="tool-section-label"&gt;Output&lt;/div&gt;
                      &lt;pre class="agent-session-pre"&gt;&lt;code&gt;13 files moved in 24 seconds. Morning note: I left 6 files alone. I filed 13 files into 3 Documents, 2 Invoices, 2 Reading. 4 of them needed the model and the rest were rules.&lt;/code&gt;&lt;/pre&gt;
                    &lt;/div&gt;
                &lt;/div&gt;
              &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;

      &lt;div class="agent-session-message agent-session-assistant"&gt;
        &lt;div class="agent-session-role-badge agent-session-role-assistant"&gt;
          Agent
        &lt;/div&gt;
        &lt;div class="agent-session-content"&gt;
                &lt;div class="agent-session-text"&gt;
                  &lt;p&gt;Shipped: 27 tests passing, dashboard screenshots from a real run, README, MIT license, render.yaml for the hosted demo. The repo is github.com/mukuvi/dozer and the Week 3 post is on DEV with the #hf26challenge tag.&lt;/p&gt;

                &lt;/div&gt;
        &lt;/div&gt;
      &lt;/div&gt;
  &lt;/div&gt;

  &lt;div class="agent-session-footer"&gt;
    &lt;span class="agent-session-meta"&gt;
        8 of 8 messages
    &lt;/span&gt;
  &lt;/div&gt;
&lt;/div&gt;


&lt;h2&gt;
  
  
  The Lazy Payoff
&lt;/h2&gt;

&lt;p&gt;Thirteen files moved without me. Three names got better. One duplicate pair got separated, and neither copy was deleted.&lt;/p&gt;

&lt;p&gt;The numbers are less flattering than a pitch. The six files left alone are six decisions I still have to make. Dozer did not remove the work. It batched it and moved it to a moment when I might feel like doing it. What it did remove was the interruption: you open the folder for one file and leave twenty minutes later with nothing decided.&lt;/p&gt;

&lt;p&gt;Machine time: 24 seconds, paced at two seconds per file. Human time: one glance at a morning note.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma&lt;/strong&gt;: Gemma 3 1B (&lt;code&gt;Q4_K_M&lt;/code&gt;, 806 MB) does every ambiguous classification, running locally through &lt;code&gt;llama.cpp&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Render&lt;/strong&gt;: &lt;code&gt;render.yaml&lt;/code&gt; deploys the dashboard as a hosted demo. The free instance cannot hold a 0.8 GB model, so the hosted demo runs the rule engine only, on a throwaway copy of the sample folder, and the UI says so. The model path is the one you run at home.&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;&lt;em&gt;Built for Hacktoberfest 2026. No API key, no upload, no account, and nothing leaves the machine.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>hf26challenge</category>
      <category>devchallenge</category>
      <category>ai</category>
      <category>opensource</category>
    </item>
    <item>
      <title>Quiet Ledger: a receipt sorter that never uploads your receipts</title>
      <dc:creator>James Ngandu</dc:creator>
      <pubDate>Sun, 18 Oct 2026 17:37:15 +0000</pubDate>
      <link>https://dev.to/mukuvi/quiet-ledger-a-receipt-sorter-that-never-uploads-your-receipts-ion</link>
      <guid>https://dev.to/mukuvi/quiet-ledger-a-receipt-sorter-that-never-uploads-your-receipts-ion</guid>
      <description>&lt;p&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hf26"&gt;Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fg9hb3a22nlhd8inf0q9o.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fg9hb3a22nlhd8inf0q9o.png" alt="Quiet Ledger's ledger view" width="800" height="719"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;Quiet Ledger is a receipt and invoice sorter for one-person businesses, and it runs entirely on your own machine.&lt;/p&gt;

&lt;p&gt;You drop a PDF invoice, a photo of a receipt, or paste the text. Gemma 3 1B reads it locally, drafts a bookkeeping entry, and shows it to you. You fix anything it got wrong, and then you export a CSV with the columns an accountant actually wants.&lt;/p&gt;

&lt;p&gt;That is the whole loop. It exists because I watched a friend who freelances spend the last Sunday of every month with a shoebox of receipts, a bank statement, and a spreadsheet. The work is not hard. It is just relentless, and every cloud tool wants the receipts uploaded to a server first.&lt;/p&gt;

&lt;p&gt;Three things make it different from the receipt apps I tried:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The document never leaves the laptop.&lt;/strong&gt; No upload, no account, no API key, no per-document fee.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The model drafts; you approve.&lt;/strong&gt; This is not a black box that files your taxes. It is a fast first pass you correct in a few seconds.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The output is a file you own.&lt;/strong&gt; &lt;code&gt;data/receipts.jsonl&lt;/code&gt; on disk, and a CSV you can hand to anyone.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;It reads the things a freelancer's shoebox actually contains: supermarket receipts, a hardware store tax invoice, a design retainer from another country. It keeps currencies separate instead of quietly adding shillings to dollars, which is the kind of bug that ruins a month.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;The demo is three commands, because the whole point is that it runs where the documents are:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;make &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; make models &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; make llama
&lt;span class="nb"&gt;export &lt;/span&gt;&lt;span class="nv"&gt;LLAMA_BIN&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$PWD&lt;/span&gt;&lt;span class="s2"&gt;/.llama.cpp/build/bin/llama-cli"&lt;/span&gt;
make run            &lt;span class="c"&gt;# http://127.0.0.1:8000&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;There is no hosted demo link, and that is deliberate: hosting it would mean uploading other people's receipts, which is the exact thing this project refuses to do. The repo ships three sample documents so the screenshot above is reproducible from a clean clone.&lt;/p&gt;

&lt;p&gt;What a real run looks like. I dropped the sample tax invoice PDF, and 43 seconds later on four CPU cores this came back:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"date"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"2026-10-12"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"vendor"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"KASHA HARDWARE LTD"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"category"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"hardware"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"total"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mf"&gt;4850.0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"currency"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"KES"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"tax"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mf"&gt;620.0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"payment_method"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"mpesa"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"invoice_number"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"INV-2026-1044"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"extracted_by"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"gemma-3-1b-it"&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;And the export:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;date,vendor,category,total,currency,tax,payment_method,invoice_number
2026-10-15,MAMA NJERI'S KITCHEN,meals,1400.00,KES,,cash,
2026-10-12,KASHA HARDWARE LTD,hardware,4850.00,KES,620.00,mpesa,INV-2026-1044
2026-10-09,Northwind Studio LLC,professional-services,2083.20,USD,163.20,bank transfer,NW-8841
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://github.com/mukuvi/quiet-ledger" rel="noopener noreferrer"&gt;https://github.com/mukuvi/quiet-ledger&lt;/a&gt; (MIT licensed).&lt;/p&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;A small FastAPI server, a single-page PWA, and the plumbing from my last offline project, collapsed into one app: Gemma 3 1B with &lt;code&gt;llama.cpp&lt;/code&gt;, plus &lt;code&gt;pdftotext&lt;/code&gt; and &lt;code&gt;tesseract&lt;/code&gt; for getting text out of files.&lt;br&gt;
&lt;/p&gt;

&lt;pre data-lang="mermaid"&gt;&lt;code&gt;flowchart LR
  A[PDF / photo / paste] --&amp;gt; B[FastAPI]
  B --&amp;gt; C[pdftotext or tesseract]
  C --&amp;gt; D[Gemma 3 1B via llama.cpp]
  D --&amp;gt; E[normalise + sanity checks]
  E --&amp;gt; F[you approve or fix]
  F --&amp;gt; G[(data/receipts.jsonl)]
  G --&amp;gt; H[CSV for the accountant]&lt;/code&gt;&lt;/pre&gt;



&lt;p&gt;The model gets a strict, JSON-only prompt at temperature 0. No chat, no markdown, no commentary, and an explicit instruction to return &lt;code&gt;null&lt;/code&gt; rather than invent a value:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;- Return JSON only. No markdown, no explanation, no commentary.
- Use null when a value is not printed on the receipt. Never invent values.
- "total" is the final amount paid. If the receipt shows both a subtotal
  and a total or amount due, use the final amount due, never the subtotal.
- "tax" is the tax or VAT amount in money, not the tax rate.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That last pair of rules exists because I watched the model get both wrong.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The normaliser is where the real work is.&lt;/strong&gt; A 1B model is a good reader and a mediocre accountant. Three failure modes showed up in my own test runs, and each one is now handled in code rather than papered over:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;It returned &lt;code&gt;16%&lt;/code&gt;, the VAT rate, in the &lt;code&gt;tax&lt;/code&gt; field. Money and rates are now parsed differently, and a tax amount smaller than one percent of the total is treated as a rate and replaced with the printed tax line.&lt;/li&gt;
&lt;li&gt;It returned the &lt;strong&gt;subtotal&lt;/strong&gt; as the total on an invoice that also had an &lt;code&gt;Amount due&lt;/code&gt; line. The printed amount due now wins over the model's answer.&lt;/li&gt;
&lt;li&gt;It labelled a hardware purchase &lt;code&gt;other&lt;/code&gt;. When the model has no useful opinion, a keyword pass over the receipt text gets a second chance, so &lt;code&gt;hardware&lt;/code&gt; wins over &lt;code&gt;other&lt;/code&gt;.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Everything else is deliberately boring. Amounts lose their symbols and separators, dates become ISO, categories map onto a fixed list of twelve, and a payment method of &lt;code&gt;M-PESA&lt;/code&gt; becomes &lt;code&gt;mpesa&lt;/code&gt; so it does not split into three spellings. Files are deduplicated by SHA-256, so dragging the same receipt in twice does nothing.&lt;/p&gt;

&lt;p&gt;Storage is JSON Lines. Not Postgres, not SQLite, not a vector database. A freelancer has a few thousand receipts, not a data warehouse, and a plain file is easier to back up, grep, and explain.&lt;/p&gt;

&lt;p&gt;If no model is on disk, the app still runs and labels every entry &lt;code&gt;fallback-parser&lt;/code&gt;. I did that on purpose: someone should be able to clone the repo, see the workflow, and decide whether the model is worth the download before committing to it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who Would Pay For This
&lt;/h2&gt;

&lt;p&gt;The person I built it for is a freelancer or a two-person studio with client receipts in three currencies and an accountant who charges by the hour to sort them.&lt;/p&gt;

&lt;p&gt;The pitch is not "another subscription". It is:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Privacy that is real, not a policy page.&lt;/strong&gt; Client names, supplier prices, and where you were on a Tuesday never leave the machine. For anyone under a confidentiality clause, that is the product.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Zero marginal cost.&lt;/strong&gt; No per-receipt pricing and no seat count. The download is 0.8 GB, once.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It works where the work happens.&lt;/strong&gt; A market trader with a laptop and no reliable connection can sort receipts in the evening. A cloud tool cannot.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The accountant gets the same CSV either way.&lt;/strong&gt; The value is the hour back, not a new format to learn.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The honest price for this is low, and that is fine. A tool that costs nothing to run can be priced at what the saved time is worth, and the model can be swapped for a bigger one later without renegotiating a contract.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;Every part of this that matters is open, and each one is a practical advantage rather than a slogan.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The document stays private because inference is local.&lt;/strong&gt; A receipt is one of the most sensitive documents a small business owns. A closed sorter has to move it to a server to read it. Running the weights on the same laptop means there is no trust boundary to audit, because there is no boundary.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It runs with no network.&lt;/strong&gt; The model, the prompt, and the parser are all on disk. The app is a PWA that opens offline. In a place where connectivity is metered or missing, the open version is not a worse version of the cloud one. It is the only version that works.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;I can change the accountant.&lt;/strong&gt; Gemma runs from a local GGUF file, so moving to a larger Gemma, a different quantisation, or a different model entirely is an environment variable, not a vendor conversation. A regional list of vendor names and tax labels is a fine-tuning run away. That kind of adaptation is impossible when the model is someone else's endpoint.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;I can inspect the reasoning.&lt;/strong&gt; When the model put a tax rate in the tax field, I could read the prompt, change it, and re-run the same document. With a closed API I would be guessing at what changed between two versions.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The cost stays at zero.&lt;/strong&gt; No API key, no rate limit, no bill that grows with how diligent you are about keeping receipts. That matters most for the people who most need the tool.&lt;/p&gt;

&lt;p&gt;The open pieces are not a licensing detail here. They are the reason the product can promise privacy, offline use, and a price of nothing.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma&lt;/strong&gt;: Gemma 3 1B runs locally through &lt;code&gt;llama.cpp&lt;/code&gt; and does every extraction in the app. Its output is then validated by the normaliser, which is the part that makes a 1B model trustworthy enough for bookkeeping.&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;&lt;em&gt;Built for Hacktoberfest 2026. No API key, no upload, no outbound calls when you sort a receipt.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>hf26challenge</category>
      <category>devchallenge</category>
      <category>ai</category>
      <category>opensource</category>
    </item>
  </channel>
</rss>
