<?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: NIDHAN JAIN</title>
    <description>The latest articles on DEV Community by NIDHAN JAIN (@nidhan_jain).</description>
    <link>https://dev.to/nidhan_jain</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%2F3713003%2F47b67f2d-37c1-4528-879f-0f278c219bbd.png</url>
      <title>DEV Community: NIDHAN JAIN</title>
      <link>https://dev.to/nidhan_jain</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/nidhan_jain"/>
    <language>en</language>
    <item>
      <title>THE UNCLAIMED: a digital archaeology field kit</title>
      <dc:creator>NIDHAN JAIN</dc:creator>
      <pubDate>Fri, 30 Oct 2026 19:58:30 +0000</pubDate>
      <link>https://dev.to/nidhan_jain/the-unclaimed-a-digital-archaeology-field-kit-142m</link>
      <guid>https://dev.to/nidhan_jain/the-unclaimed-a-digital-archaeology-field-kit-142m</guid>
      <description>&lt;p&gt;&lt;em&gt;This is my 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;/em&gt;&lt;/p&gt;

&lt;h1&gt;
  
  
  THE UNCLAIMED: a digital archaeology field kit
&lt;/h1&gt;

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

&lt;p&gt;&lt;strong&gt;THE UNCLAIMED&lt;/strong&gt; is a workspace for investigating what remains of forgotten websites. Each investigation is a case: gather search leads, inspect a live page, look for historical captures, save observations, and connect the evidence in a timeline, graph, and report.&lt;/p&gt;

&lt;p&gt;The interface is intentionally eerie. The evidence trail is deliberately ordinary: a search snippet is only a lead, and a finding is grounded in text saved to the case.&lt;/p&gt;

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

&lt;p&gt;Try the deployed app at &lt;a href="https://unclaimed.nidhan.tech/" rel="noopener noreferrer"&gt;unclaimed.nidhan.tech&lt;/a&gt;.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://github.com/Dev-Nidhan/the-unclaimed" rel="noopener noreferrer"&gt;Source repository&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://github.com/Dev-Nidhan/the-unclaimed/blob/main/docs/architecture.svg" rel="noopener noreferrer"&gt;Production architecture diagram&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://github.com/Dev-Nidhan/the-unclaimed/tree/main/docs/media" rel="noopener noreferrer"&gt;Screenshots and walkthrough files&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://github.com/Dev-Nidhan/the-unclaimed/blob/main/DEPLOYMENT.md" rel="noopener noreferrer"&gt;Deployment and environment setup&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The repository is public, but it currently has no license file. The code is visible; reuse rights have not been granted.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I built it
&lt;/h2&gt;

&lt;p&gt;The app uses Next.js, TypeScript, Drizzle, and Tiger Data Postgres. SerpApi turns searches into leads, the Internet Archive provides historical capture records, and Browserbase runs temporary remote Chromium sessions for inspecting live pages.&lt;/p&gt;

&lt;p&gt;For analysis, the server sends Gemma 3 4B a bounded packet of saved case evidence. The model returns structured findings that are checked against a schema. A claim marked verified must quote text already stored in the investigation. The graph, timeline, and report are drawn from those saved records.&lt;/p&gt;

&lt;p&gt;The model is Gemma 3 4B in both environments. On a laptop, Ollama runs locally. On Render, the Next.js app and model run as separate services; the model service is private and reached through an authenticated gateway. The Ollama port is not exposed publicly, and investigation prompts are not sent to a cloud LLM.&lt;/p&gt;

&lt;p&gt;Ambient and effect sounds are static MP3 files in the repository. A local utility uses ElevenLabs to generate the clips; the deployed app serves the files and does not call ElevenLabs at runtime.&lt;/p&gt;

&lt;h2&gt;
  
  
  Run it locally
&lt;/h2&gt;

&lt;p&gt;You need Node.js, Ollama with Gemma 3 4B, and a PostgreSQL database. Search and remote-browser features also need their provider credentials. See the linked setup guide for the full environment-variable list.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/Dev-Nidhan/the-unclaimed.git
&lt;span class="nb"&gt;cd &lt;/span&gt;the-unclaimed
npm &lt;span class="nb"&gt;install
cp&lt;/span&gt; .env.example .env.local
ollama pull gemma3:4b
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Fill in the required values in &lt;code&gt;.env.local&lt;/code&gt;, start Ollama, then run:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm run db:migrate
npm run dev
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Open &lt;code&gt;http://localhost:3000&lt;/code&gt;. The model URL defaults to &lt;code&gt;http://127.0.0.1:11434&lt;/code&gt;; the app rejects non-loopback Ollama URLs for local inference.&lt;/p&gt;

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

&lt;p&gt;A forgotten page is already easy to misremember. I wanted the AI layer to make claims that can be traced back to stored evidence, with the same Gemma model available locally for inspection and experimentation. Keeping the model self-hosted also means the core analysis path does not depend on a per-token cloud LLM API.&lt;/p&gt;

&lt;p&gt;The whole product is not offline: discovery, archived pages, remote browser sessions, and the hosted database use network services. On Render, the model is kept private while the rest of the demo remains easy to try.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Overall — Hacktoberfest Open-Source AI Challenge Week 4:&lt;/strong&gt; a digital archaeology project about investigating the traces left by vanished websites.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma:&lt;/strong&gt; Gemma 3 4B analyzes the saved evidence packet and returns structured findings.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Render:&lt;/strong&gt; Render hosts the public Next.js app and the separate private Gemma/Ollama service.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Browserbase:&lt;/strong&gt; Browserbase provides temporary remote Chromium sessions for fieldwork.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of SerpApi:&lt;/strong&gt; Search results seed the case as leads to verify.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Tiger Data:&lt;/strong&gt; Tiger Data Postgres stores investigations, sources, observations, and findings.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Sentry Agent Tracing:&lt;/strong&gt; Sentry receives scrubbed server errors and traces.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of ElevenLabs:&lt;/strong&gt; ElevenLabs generates the sound assets through a local utility; playback uses the checked-in static files.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of a .Tech Domain:&lt;/strong&gt; The live project is hosted at &lt;a href="https://unclaimed.nidhan.tech/" rel="noopener noreferrer"&gt;unclaimed.nidhan.tech&lt;/a&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;DigitalOcean Spaces is supported in the code for production artifact storage, but it is not configured for this demo, so I have not listed it as a category used.&lt;/p&gt;

&lt;h2&gt;
  
  
  Limits
&lt;/h2&gt;

&lt;p&gt;Internet Archive lookups can be rate-limited, and search results can be incomplete. Gemma can still make mistakes; evidence quoting and schema checks help make findings reviewable, but they do not replace human judgment. Production artifact uploads are not enabled without Spaces credentials.&lt;/p&gt;

&lt;p&gt;The goal is not to make the model tell a ghost story. It is to give an investigator a clear path from a surviving trace to a claim they can check.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>gemma</category>
      <category>halloween</category>
    </item>
    <item>
      <title>ClassSloth: a local Gemma planner for the week's homework</title>
      <dc:creator>NIDHAN JAIN</dc:creator>
      <pubDate>Sun, 25 Oct 2026 20:46:57 +0000</pubDate>
      <link>https://dev.to/nidhan_jain/classsloth-a-local-gemma-planner-for-the-weeks-homework-1ahc</link>
      <guid>https://dev.to/nidhan_jain/classsloth-a-local-gemma-planner-for-the-weeks-homework-1ahc</guid>
      <description>&lt;p&gt;&lt;em&gt;This is my submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19"&gt;Hacktoberfest Open-Source AI Challenge: Week 3&lt;/a&gt;, Slow and Steady. ClassSloth is a new project built for this round.&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;ClassSloth&lt;/strong&gt; answers the four questions I otherwise answer by opening every Classroom tab: what is due, what to do first, where the material is, and how to revise from notes I already have.&lt;/p&gt;

&lt;p&gt;The lazy payoff is the tab-checking. The dashboard puts one next task on the screen. Due dates, titles, and links come from Classroom or from a clearly labeled demo set. If the app cannot tell whether something was turned in, the status stays &lt;strong&gt;unknown&lt;/strong&gt;. Gemma's priorities and session lengths are estimates, and the source due date stays on the assignment.&lt;/p&gt;

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

&lt;p&gt;About 63 seconds, recorded from the running app on a laptop. It uses the labeled demo week, a saved Gemma plan, course links, and a local Ollama check. The file in the repository is &lt;code&gt;docs/demo.webm&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/Dev-Nidhan/classsloth/blob/main/docs/demo.webm" rel="noopener noreferrer"&gt;Watch the demo&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The hosted copy is &lt;a href="https://classsloth.onrender.com" rel="noopener noreferrer"&gt;https://classsloth.onrender.com&lt;/a&gt;. Gemma 3 4B runs in that Render container for the labeled assignments and one bundled note. Real Classroom accounts and private uploads are not accepted there. The first plan after a cold start waits while the weights finish downloading.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://github.com/Dev-Nidhan/classsloth" rel="noopener noreferrer"&gt;Source&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://github.com/Dev-Nidhan/classsloth#setup" rel="noopener noreferrer"&gt;Setup, OAuth, offline limits, and license&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://github.com/Dev-Nidhan/classsloth/blob/main/render.yaml" rel="noopener noreferrer"&gt;Render blueprint&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;License: MIT, file &lt;code&gt;LICENSE&lt;/code&gt; in the repository.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How I built it
&lt;/h2&gt;

&lt;p&gt;Python, Streamlit, SQLite, PyMuPDF, and Ollama. The default model is Gemma 3 4B (&lt;code&gt;gemma3:4b&lt;/code&gt;). Google Classroom is optional and read-only: courses, published coursework, course materials, and the student's own submission state. The app does not create, grade, or turn in work.&lt;br&gt;
&lt;/p&gt;

&lt;pre data-lang="mermaid"&gt;&lt;code&gt;flowchart LR
    Student[Student] --&amp;gt; UI[Streamlit]
    UI --&amp;gt; DB[(SQLite)]
    UI --&amp;gt;|Only when you click plan, match, or revise| Gemma[Ollama + Gemma 3 4B]
    Classroom[Google Classroom read-only] --&amp;gt;|Sync needs internet| DB
    Notes[Uploaded PDF or text] --&amp;gt; DB
    Gemma --&amp;gt;|Validated JSON| UI&lt;/code&gt;&lt;/pre&gt;



&lt;p&gt;On a laptop, Ollama is the local daemon. On Render, the Docker image starts Ollama on &lt;code&gt;127.0.0.1:11434&lt;/code&gt;, pulls &lt;code&gt;gemma3:4b&lt;/code&gt; onto a Render disk, and publishes only the Streamlit port. The instance is the 8 GB plan because the 4B weights are about 3.3 GB and need room to run. A $50 credit covers roughly a week at that size.&lt;/p&gt;

&lt;p&gt;Inference runs only when you ask for a plan, a note match, a revision pack, or a search summary. A Streamlit rerun does not call the model again. Assignment text and uploaded notes are wrapped as data, not instructions.&lt;/p&gt;

&lt;p&gt;Gemma 3 4B will sometimes return a valid plan shape with an empty &lt;code&gt;items&lt;/code&gt; list and the real work sitting in &lt;code&gt;study_sessions&lt;/code&gt;. ClassSloth keeps those session notes as the steps and still shows the due date from SQLite. Unknown page numbers are dropped. That matches what builders report about small local models and structured JSON: the schema helps, and you still validate. See &lt;a href="https://dev.to/jangwook_kim_e31e7291ad98/ollama-structured-outputs-in-practice-getting-type-safe-json-from-local-llms-with-pydantic-m38"&gt;Ollama structured outputs with Pydantic&lt;/a&gt; and this &lt;a href="https://dev.to/sandeep_chakravartty_219e/build-a-production-grade-ai-dashboard-architectural-insights-from-the-bmw-autotrend-platform-4hg2"&gt;local Ollama plus SQLite dashboard&lt;/a&gt;, which keeps a non-model path when the model is down. A 4B model also needs several gigabytes of RAM even when the file looks small; &lt;a href="https://dev.to/doogal/run-llms-locally-ollama-setup-hardware-requirements-4deh"&gt;this Ollama hardware note&lt;/a&gt; is why the Render service is not on the free 512 MB instance.&lt;/p&gt;

&lt;p&gt;The test suite does not need Google, SerpApi, or ElevenLabs credentials. The last run here was &lt;strong&gt;41 passed&lt;/strong&gt;.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="n"&gt;git&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;clone&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;https://github.com/Dev-Nidhan/classsloth.git&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="n"&gt;cd&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;classsloth&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="n"&gt;python&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nt"&gt;-m&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;venv&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;venv&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;\.venv\Scripts\Activate.ps1&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="n"&gt;python&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nt"&gt;-m&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;pip&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;install&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nt"&gt;-r&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;requirements.txt&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="n"&gt;Copy-Item&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;env&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;example&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;env&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="nx"&gt;ollama&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;pull&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;gemma3:4b&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="n"&gt;streamlit&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;run&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;app.py&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;With no OAuth client, choose &lt;strong&gt;Load demo week&lt;/strong&gt;. Every synthetic title starts with &lt;code&gt;Demo:&lt;/code&gt;.&lt;/p&gt;

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

&lt;p&gt;Coursework and notes are personal. On your own computer, the plan, the PDF text, and the revision pack stay on the machine. The public Render copy uses the labeled demo week and one bundled note, not a student's Classroom account. There is no analytics and no separate public model port.&lt;/p&gt;

&lt;p&gt;Running Gemma yourself also means the weekly plan does not depend on a per-token API bill. It does depend on memory. On a laptop a plan came back in well under a minute, and one run omitted the per-assignment steps until the validator repaired them from the session notes Gemma did return.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Overall&lt;/strong&gt;, Hacktoberfest Open-Source AI Challenge: Week 3, Slow and Steady.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma.&lt;/strong&gt; Gemma 3 4B is the planner, the note matcher, and the exam-pack writer, both on a laptop and inside the Render container.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Render.&lt;/strong&gt; The live service is &lt;a href="https://classsloth.onrender.com" rel="noopener noreferrer"&gt;classsloth.onrender.com&lt;/a&gt;, a Docker web service on an 8 GB instance. Ollama and &lt;code&gt;gemma3:4b&lt;/code&gt; stay on localhost, and the weights sit on a Render disk. Visitors only reach Streamlit. Google login and private uploads are disabled there.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of SerpApi.&lt;/strong&gt; A topic search keeps the titles, URLs, and snippets SerpApi actually returned, prefers known educational hosts, and asks Gemma to order only those URLs. Invented links are dropped. It runs only from a button and needs &lt;code&gt;SERPAPI_API_KEY&lt;/code&gt;. The public demo leaves this off.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of ElevenLabs.&lt;/strong&gt; Exam Prep speaks a passage you select, after a checkbox that says the text leaves the computer. Passages over 2,500 characters are refused before any request. It needs &lt;code&gt;ELEVENLABS_API_KEY&lt;/code&gt;. The public demo leaves this off.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>gemma</category>
      <category>python</category>
    </item>
    <item>
      <title>RepairDesk Local: A Private Repair Desk with Optional Gemma AI</title>
      <dc:creator>NIDHAN JAIN</dc:creator>
      <pubDate>Sun, 18 Oct 2026 21:17:57 +0000</pubDate>
      <link>https://dev.to/nidhan_jain/repairdesk-local-a-private-repair-desk-with-optional-gemma-ai-5614</link>
      <guid>https://dev.to/nidhan_jain/repairdesk-local-a-private-repair-desk-with-optional-gemma-ai-5614</guid>
      <description>&lt;p&gt;&lt;em&gt;This is my submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week2-2026-10-12"&gt;Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle&lt;/a&gt;. I started this new project during the Week 2 challenge window.&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;RepairDesk Local&lt;/strong&gt; is an offline-first repair desk for independent phone and laptop repair shops. It tracks jobs from intake through completion, calculates itemized estimates, and prepares customer-update drafts. The core workflow is ordinary Python backed by SQLite; AI is optional.&lt;/p&gt;

&lt;p&gt;The intended user is a small repair shop or independent technician who wants a lightweight way to organize repairs without sending records to a hosted repair SaaS. I have not yet validated willingness to pay with a shop or claimed paying users. The goal is a focused tool that a shop could install and operate locally, with no per-request AI API charge.&lt;/p&gt;

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

&lt;p&gt;Demo Video:&lt;br&gt;
&lt;/p&gt;
&lt;div class="crayons-card c-embed text-styles text-styles--secondary"&gt;
    &lt;div class="c-embed__content"&gt;
      &lt;div class="c-embed__body"&gt;
        &lt;h2 class="fs-xl lh-tight"&gt;
          &lt;a href="https://drive.google.com/file/d/1MORm65X9iEvioMJyzaj_oIuGq5mWv338/view?usp=sharing" rel="noopener noreferrer" class="c-link"&gt;
            RepairDesk Local Demo.mp4 - Google Drive
          &lt;/a&gt;
        &lt;/h2&gt;
        &lt;div class="color-secondary fs-s flex items-center"&gt;
            &lt;img alt="favicon" class="c-embed__favicon m-0 mr-2 radius-0" src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fssl.gstatic.com%2Fdocs%2Fdoclist%2Fimages%2Fdrive_favicon_2026_32dp.png" width="32" height="32"&gt;
          drive.google.com
        &lt;/div&gt;
      &lt;/div&gt;
    &lt;/div&gt;
&lt;/div&gt;


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

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://github.com/Dev-Nidhan/repairdesk-local" rel="noopener noreferrer"&gt;Source code&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://github.com/Dev-Nidhan/repairdesk-local#installation" rel="noopener noreferrer"&gt;Installation, configuration, and architecture&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How I built it
&lt;/h2&gt;

&lt;p&gt;The stack is Python, Streamlit, SQLite, Pydantic, and Ollama. The default optional model is Gemma 3 4B (&lt;code&gt;gemma3:4b&lt;/code&gt;). The regular job workflow and estimate arithmetic do not depend on a model. When requested, Gemma produces an intake checklist and possible explanations; Pydantic validates the structured response, and suggestions remain explicitly unconfirmed. The model cannot set prices, approve work, or send customer updates.&lt;br&gt;
&lt;/p&gt;

&lt;pre data-lang="mermaid"&gt;&lt;code&gt;flowchart LR
    Staff[Shop staff] --&amp;gt; UI[Streamlit on local computer]
    UI --&amp;gt; DB[(SQLite repair database)]
    UI --&amp;gt;|Optional device and complaint context; no customer name/reference| AI[Python AI service]
    AI --&amp;gt;|Default: localhost HTTP| Ollama[Ollama + Gemma 3 4B]
    Ollama --&amp;gt; AI
    AI --&amp;gt;|Validated, unconfirmed suggestions| UI
    UI --&amp;gt;|Human reviews and copies manually| Customer[Customer]&lt;/code&gt;&lt;/pre&gt;



&lt;p&gt;With the default &lt;code&gt;OLLAMA_BASE_URL=http://localhost:11434&lt;/code&gt;, the model request goes to Ollama on the same computer. This application has no cloud AI integration. If an operator deliberately configures a remote Ollama endpoint, the selected device and complaint context will be sent to that endpoint instead. Customer names and references are omitted from model prompts. The app warns against saving passwords, PINs, or device unlock codes.&lt;/p&gt;

&lt;h3&gt;
  
  
  Try it locally
&lt;/h3&gt;

&lt;p&gt;Python 3.11 or newer is required. From PowerShell:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="n"&gt;git&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;clone&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;https://github.com/Dev-Nidhan/repairdesk-local.git&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="n"&gt;cd&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;repairdesk-local&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="n"&gt;python&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nt"&gt;-m&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;venv&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;venv&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;\.venv\Scripts\Activate.ps1&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="n"&gt;python&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nt"&gt;-m&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;pip&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;install&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nt"&gt;-r&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;requirements.txt&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="n"&gt;Copy-Item&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;env&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;example&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;env&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="nx"&gt;ollama&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;pull&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;gemma3:4b&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="n"&gt;streamlit&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;run&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;app.py&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Ollama is considered to be installed through &lt;a href="https://ollama.com/download" rel="noopener noreferrer"&gt;Ollama download&lt;/a&gt; before pulling gemma3:4b model, though Ollama and Gemma are optional; the job desk, estimates, and template-based updates still work without them. The test suite currently has 30 passing tests and does not call a live model.&lt;/p&gt;

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

&lt;p&gt;Repair complaints can contain sensitive details. Keeping SQLite and, by default, Ollama on the shop's own computer gives the operator control over that data and allows the essential workflow to work without a hosted AI dependency. An open-weight model also avoids a per-prompt API bill.&lt;/p&gt;

&lt;p&gt;That does not make local inference free or equally fast on every device: operators provide the hardware and electricity, and CPU performance and model quality vary. I am treating Gemma as a reviewable helper, not a substitute for a technician or a claim that every shop will save money. For more on the trade-offs, see &lt;a href="https://dev.to/dekita/i-built-an-offline-invoice-sorter-for-freelancers-then-measured-whether-local-inference-is-35p4"&gt;a builder's measured cost comparison of an offline invoice sorter&lt;/a&gt; and this &lt;a href="https://dev.to/doogal/run-llms-locally-ollama-setup-hardware-requirements-4deh"&gt;Ollama setup and hardware guide&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  First user and validation
&lt;/h2&gt;

&lt;p&gt;I have not yet tested this with an external repair shop or collected revenue. The walkthrough uses only fictional sample records. The next validation step is to ask independent technicians whether local job tracking, estimate preparation, and optional intake suggestions save enough time to be useful.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Overall Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle.&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma:&lt;/strong&gt; Gemma 3 4B is an implemented, optional model for structured repair-intake suggestions through Ollama; this is a local workflow, not a model name added only for the submission.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of GitHub Copilot:&lt;/strong&gt; I used Copilot in VS Code to help build and iterate on the application and its documentation. The shipped app does not call Copilot at runtime.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>gemma</category>
      <category>python</category>
    </item>
  </channel>
</rss>
