<?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: Kaushani Koner</title>
    <description>The latest articles on DEV Community by Kaushani Koner (@kaushanikoner20).</description>
    <link>https://dev.to/kaushanikoner20</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%2F4158397%2F56e501ee-63f8-43c9-85e1-94c142ee82ff.jpeg</url>
      <title>DEV Community: Kaushani Koner</title>
      <link>https://dev.to/kaushanikoner20</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/kaushanikoner20"/>
    <language>en</language>
    <item>
      <title>Unearthing the Ghosts of the Early Internet with Local AI 🕯️</title>
      <dc:creator>Kaushani Koner</dc:creator>
      <pubDate>Fri, 30 Oct 2026 17:59:54 +0000</pubDate>
      <link>https://dev.to/kaushanikoner20/unearthing-the-ghosts-of-the-early-internet-with-local-ai-5hcg</link>
      <guid>https://dev.to/kaushanikoner20/unearthing-the-ghosts-of-the-early-internet-with-local-ai-5hcg</guid>
      <description>&lt;h2&gt;
  
  
  The Internet Left a Trail. I Built an AI Séance to Follow It. 🕯️
&lt;/h2&gt;

&lt;p&gt;What if we could investigate the forgotten corners of the early internet—not through nostalgia or guesswork, but through the evidence that archived webpages actually left behind?&lt;/p&gt;

&lt;p&gt;That question inspired &lt;strong&gt;Web Séance&lt;/strong&gt;, my project for Hacktoberfest 2026 Week 4: &lt;em&gt;Ghost in the Machine&lt;/em&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  🎬 A Glimpse Into Web Séance
&lt;/h2&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%2Fzpb61yagjs269gzbfekx.gif" 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%2Fzpb61yagjs269gzbfekx.gif" alt="demo video-gif of the project" width="800" height="495"&gt;&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/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F8ljuhwj2aw2oxsm20wqd.gif" 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%2F8ljuhwj2aw2oxsm20wqd.gif" alt="real application demo" width="600" height="272"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;This animated preview walks through the project's workflow. It offers a visual walkthrough of the experience; the full application and its source code are available below.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Interactive demo walkthrough:&lt;/strong&gt; &lt;a href="https://kaushanikoner20-hub.github.io/Web-Seance/demo.html" rel="noopener noreferrer"&gt;https://kaushanikoner20-hub.github.io/Web-Seance/demo.html&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;GitHub repository:&lt;/strong&gt; &lt;a href="https://github.com/kaushanikoner20-hub/Web-Seance" rel="noopener noreferrer"&gt;https://github.com/kaushanikoner20-hub/Web-Seance&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

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

&lt;p&gt;Web Séance is an AI-assisted digital archaeology tool for exploring forgotten internet culture through historical webpages preserved by the Internet Archive.&lt;/p&gt;

&lt;p&gt;Web Séance is designed for digital historians, web archivists, researchers, and curious developers who want to explore forgotten internet culture through verifiable archival evidence.&lt;/p&gt;

&lt;p&gt;The application helps users:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Retrieve historical webpage captures using the Internet Archive's CDX API.&lt;/li&gt;
&lt;li&gt;Extract outgoing hyperlinks from archived HTML.&lt;/li&gt;
&lt;li&gt;Normalize and classify links.&lt;/li&gt;
&lt;li&gt;Assign traceable Evidence IDs to extracted links.&lt;/li&gt;
&lt;li&gt;Explore observed hyperlink trails through a D3.js network graph.&lt;/li&gt;
&lt;li&gt;Ask a locally running Gemma model to interpret the collected evidence.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The goal isn't to manufacture a story about the past. It's to make the available evidence easier to inspect.&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%2Focgbvf6ev312m66znabb.gif" 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%2Focgbvf6ev312m66znabb.gif" alt="features-flow" width="200" height="133"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  🔬 How It Works
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;1. Retrieve the archive&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The application uses the Internet Archive CDX API to locate historical captures and retrieve archived HTML.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Extract evidence&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;BeautifulSoup is used to parse archived HTML and extract outgoing hyperlinks. The extracted links receive identifiers so they can be traced back to their source page.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Ask the local AI historian&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Gemma 3:4b, running locally through Ollama, helps interpret the evidence. The system checks whether cited Evidence IDs exist in the collected evidence set.&lt;/p&gt;

&lt;p&gt;That check is useful, but it is not a guarantee that every AI-generated interpretation is factually correct. The distinction between evidence, hypotheses, and unknowns remains important.&lt;/p&gt;

&lt;h2&gt;
  
  
  🔎 A Real Investigation: Excavating a 1997 Website
&lt;/h2&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%2Fvnlritj371f9ny9biwre.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%2Fvnlritj371f9ny9biwre.png" alt="Investigation" width="799" height="359"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;To test Web Séance against real historical material, I investigated an archived webpage:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Website:&lt;/strong&gt; Jeffrey Zeldman Presents&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Original URL:&lt;/strong&gt; &lt;code&gt;http://www.zeldman.com/toc.html&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Capture date:&lt;/strong&gt; March 30, 1997&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Capture timestamp:&lt;/strong&gt; &lt;code&gt;19970330110648&lt;/code&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Web Séance retrieved the archived HTML, recorded its SHA-256 fingerprint, and extracted &lt;strong&gt;41 outgoing hyperlinks&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;Each extracted link received a unique Evidence ID, making individual findings traceable.&lt;/p&gt;

&lt;h3&gt;
  
  
  What a Hyperlink Actually Tells Us
&lt;/h3&gt;

&lt;p&gt;A hyperlink is evidence of a connection between two web pages, but it does not automatically reveal the relationship between the people behind them.&lt;/p&gt;

&lt;p&gt;For example, if an archived page links to another website, we can verify that the link existed in that capture. However, we cannot conclude that the website owners collaborated, knew each other, endorsed one another or shared a common purpose.&lt;/p&gt;

&lt;p&gt;Web Séance therefore distinguishes between &lt;strong&gt;observable link connections&lt;/strong&gt; and &lt;strong&gt;unverified interpretations&lt;/strong&gt;. This distinction is important when reconstructing the early internet, where a network of hyperlinks can reveal how pages were connected without necessarily explaining why those connections existed.&lt;/p&gt;

&lt;h3&gt;
  
  
  What did the page link to?
&lt;/h3&gt;

&lt;p&gt;To make the extracted evidence more concrete, here are five representative records from the 41 outgoing hyperlinks:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Evidence ID&lt;/th&gt;
&lt;th&gt;Anchor text&lt;/th&gt;
&lt;th&gt;Destination&lt;/th&gt;
&lt;th&gt;Scope&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;P6-L1&lt;/td&gt;
&lt;td&gt;GOOD EVENING,&lt;/td&gt;
&lt;td&gt;&lt;code&gt;/snds/bell.wav&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Same-site&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;P6-L2&lt;/td&gt;
&lt;td&gt;entrance,&lt;/td&gt;
&lt;td&gt;&lt;code&gt;/welcome.html&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Same-site&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;P6-L3&lt;/td&gt;
&lt;td&gt;—&lt;/td&gt;
&lt;td&gt;&lt;code&gt;/awards.html&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Same-site&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;P6-L4&lt;/td&gt;
&lt;td&gt;Ad Graveyard&lt;/td&gt;
&lt;td&gt;&lt;code&gt;/ad.html&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Same-site&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;P6-L5&lt;/td&gt;
&lt;td&gt;Pardon My Icons&lt;/td&gt;
&lt;td&gt;&lt;code&gt;/icon.html&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Same-site&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;em&gt;These five entries are representative examples from the 41 outgoing hyperlinks extracted from the archived page. The Evidence IDs identify individual records in the collected evidence set. Destinations are shortened for readability; the original URLs were extracted from the archived HTML.&lt;/em&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  Why does this matter?
&lt;/h3&gt;

&lt;p&gt;Instead of asking an AI to invent a story about the early internet, Web Séance starts with inspectable historical evidence.&lt;/p&gt;

&lt;p&gt;The Evidence IDs connect extracted links to their archived source page, making the investigation easier to reproduce and review.&lt;/p&gt;

&lt;p&gt;That's the distinction Web Séance is designed to preserve: &lt;strong&gt;evidence first, interpretation second.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  🧠 Why Gemma?
&lt;/h2&gt;

&lt;p&gt;I used &lt;strong&gt;Gemma 3:4b through Ollama&lt;/strong&gt; as the local AI historian.&lt;/p&gt;

&lt;p&gt;I wanted the model to work with evidence collected by the application rather than inventing historical connections from vague prompts.&lt;/p&gt;

&lt;p&gt;The division of responsibilities is intentional: Python handles archival retrieval and evidence extraction, while Gemma helps interpret the resulting evidence. This allows the model's interpretations to be examined against the records collected by the application.&lt;/p&gt;

&lt;p&gt;The intended workflow separates:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Confirmed facts:&lt;/strong&gt; What the archived HTML and extracted links directly support.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hypotheses:&lt;/strong&gt; Possible interpretations that need further investigation.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Unknowns:&lt;/strong&gt; Questions the available evidence cannot answer.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  Evaluating the Local AI Historian
&lt;/h3&gt;

&lt;p&gt;I tested Gemma 3:4B using archived webpage evidence to understand how effectively a small local model could extract facts, interpret hyperlinks and distinguish verified information from hypotheses.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Evaluation approach&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;I used the evidence dossier from the 1997 website investigation and asked Gemma questions covering five areas:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Metadata extraction:&lt;/strong&gt; Could it identify the page title, original URL and archival capture date?&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hyperlink identification:&lt;/strong&gt; Could it retrieve the supplied URLs and their associated anchor text?&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Evidence interpretation:&lt;/strong&gt; Could it distinguish an observed hyperlink from an unverified relationship or historical claim?&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Uncertainty handling:&lt;/strong&gt; Could it identify what the archived evidence did not establish?&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Citation discipline:&lt;/strong&gt; Did its responses refer to actual Evidence IDs without inventing records?&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;I reviewed the responses against the collected evidence, paying attention to factual correctness, unsupported assumptions, missing information and citation accuracy. I also experimented with shorter prompts and explicit evidence constraints to observe whether more focused instructions improved response consistency.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Observed results&lt;/strong&gt;&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Evaluation area&lt;/th&gt;
&lt;th&gt;Observed result&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Historical fact extraction&lt;/td&gt;
&lt;td&gt;Successfully retrieved page title, URL and archival date&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Hyperlink identification&lt;/td&gt;
&lt;td&gt;Correctly identified the supplied URLs and anchor text&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Link classification&lt;/td&gt;
&lt;td&gt;Demonstrated understanding of SAME-SITE and EXTERNAL classifications&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Evidence interpretation&lt;/td&gt;
&lt;td&gt;Recognized the difference between hyperlink connections and proven relationships&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Historical reasoning&lt;/td&gt;
&lt;td&gt;Generated plausible hypotheses about multimedia and early web design&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Uncertainty awareness&lt;/td&gt;
&lt;td&gt;Identified missing information about linked pages and website purpose&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Evidence accuracy&lt;/td&gt;
&lt;td&gt;Needed occasional guidance to avoid unsupported details and incorrect references&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Prompt responsiveness&lt;/td&gt;
&lt;td&gt;Produced more focused and structured answers with shorter, evidence-constrained prompts&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;These observations describe behavior on this investigation's evidence set; they should not be interpreted as a general measure of model accuracy.&lt;/p&gt;

&lt;p&gt;The experiment showed that Gemma 3:4B can support local historical evidence interpretation, particularly when given structured archival data and focused instructions. It also highlighted opportunities to improve citation validation and uncertainty handling within Web Séance.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;This was a qualitative, task-based exploratory evaluation using one historical investigation, not a comprehensive accuracy benchmark.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  🧪 How I Checked the AI's Work
&lt;/h2&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%2Fc3jh9xlt12u728i86oaq.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%2Fc3jh9xlt12u728i86oaq.png" alt="evidence 1" width="800" height="366"&gt;&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%2F6irighkraxcvcjdqf7hi.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%2F6irighkraxcvcjdqf7hi.png" alt="evidence 2" width="800" height="360"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Gemma 3:4b in action:&lt;/strong&gt; The local model analyses an evidence dossier containing 41 extracted links, separates verified facts from plausible hypotheses, and identifies unknowns. Web Séance then checks whether cited Evidence IDs exist in the dossier.&lt;/p&gt;




&lt;p&gt;That gives me a basic traceability check: a citation should point to an actual extracted record, not an invented identifier.&lt;/p&gt;

&lt;p&gt;However, this is not a complete factuality evaluation. A valid Evidence ID does not automatically mean the model interpreted that evidence correctly.&lt;/p&gt;

&lt;p&gt;I treat these as separate questions:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Traceability:&lt;/strong&gt; Does the cited Evidence ID exist?&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Interpretation:&lt;/strong&gt; Does the cited evidence actually support the statement?&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Uncertainty:&lt;/strong&gt; Does the model distinguish what is known from what remains unknown?&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This is an important limitation of the current implementation, and an area I want to evaluate more rigorously.&lt;/p&gt;




&lt;h2&gt;
  
  
  👻 What the Archives Cannot Tell Us
&lt;/h2&gt;

&lt;p&gt;One of the most important lessons from building Web Séance is that an archive is not a complete record of the internet.&lt;/p&gt;

&lt;p&gt;A missing hyperlink does not necessarily mean a connection never existed. A page may have changed, an archive capture may be incomplete, or an important interaction may have happened somewhere outside the archived HTML.&lt;/p&gt;

&lt;p&gt;I don't want the AI to turn these gaps into confident historical stories.&lt;/p&gt;

&lt;p&gt;For this reason, Web Séance treats its findings as evidence to investigate rather than a complete reconstruction of internet history.&lt;/p&gt;




&lt;h2&gt;
  
  
  🛠️ Tech Stack
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Python&lt;/strong&gt; — application logic&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Streamlit&lt;/strong&gt; — user interface&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;BeautifulSoup&lt;/strong&gt; — HTML parsing and link extraction&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Requests&lt;/strong&gt; — HTTP requests&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pandas&lt;/strong&gt; — structured data handling&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;D3.js&lt;/strong&gt; — network graph visualization&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Internet Archive CDX API&lt;/strong&gt; — historical capture discovery&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ollama + Gemma 3:4b&lt;/strong&gt; — local AI interpretation&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  🌍 Why Open Innovation Matters
&lt;/h2&gt;

&lt;p&gt;The early internet grew through shared ideas, open standards, personal websites, mailing lists, and communities that built things in public.&lt;/p&gt;

&lt;p&gt;I wanted Web Séance to follow a similar philosophy.&lt;/p&gt;

&lt;p&gt;The project combines publicly accessible web archives with an open-source AI model that can run locally through Ollama.&lt;/p&gt;

&lt;p&gt;This matters for a few reasons:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Inspectable methodology:&lt;/strong&gt; The source code makes the extraction and evidence-handling process available for others to examine.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Local AI experimentation:&lt;/strong&gt; Gemma can interpret collected evidence without sending the inference request to a hosted LLM API.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Reproducibility:&lt;/strong&gt; Other developers can inspect the implementation, test the same workflow, and challenge its conclusions.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Room to improve:&lt;/strong&gt; The project can be extended with better archival sources, stronger evaluation, and more careful historical analysis.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;There is an important distinction: local AI inference does not make the entire application offline. Retrieving new historical captures still requires access to the Internet Archive.&lt;/p&gt;

&lt;p&gt;For me, open innovation is not just about making a repository public. It's about making the process open to inspection, criticism, and improvement.&lt;/p&gt;

&lt;h2&gt;
  
  
  🕯️ The Spirit of the Project
&lt;/h2&gt;

&lt;p&gt;The internet doesn't remember everything.&lt;/p&gt;

&lt;p&gt;Some pages disappear. Some links break. Some histories become difficult to reconstruct.&lt;/p&gt;

&lt;p&gt;But sometimes, a small trace remains.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Web Séance is an attempt to follow those traces responsibly—with archival evidence first, AI interpretation second, and uncertainty always visible.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;If you're interested in digital archaeology, forgotten internet culture, open-source AI, or the history of the web, I'd love for you to explore the project and share your thoughts.&lt;/p&gt;

&lt;h2&gt;
  
  
  🏆 Prize Category
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Best Use of Gemma&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Built with Gemma 3:4b running locally through Ollama.&lt;/p&gt;




&lt;p&gt;Curious about what the early internet left behind? Explore the project, try the interactive walkthrough, and let me know what you discover.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Project:&lt;/strong&gt; &lt;a href="https://github.com/kaushanikoner20-hub/Web-Seance" rel="noopener noreferrer"&gt;https://github.com/kaushanikoner20-hub/Web-Seance&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Interactive walkthrough:&lt;/strong&gt; &lt;a href="https://kaushanikoner20-hub.github.io/Web-Seance/demo.html" rel="noopener noreferrer"&gt;https://kaushanikoner20-hub.github.io/Web-Seance/demo.html&lt;/a&gt;&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>ai</category>
      <category>opensource</category>
    </item>
    <item>
      <title>ssorg.dev: I Built a Local AI Librarian for All Those Screenshots I Keep Forgetting About 🦥</title>
      <dc:creator>Kaushani Koner</dc:creator>
      <pubDate>Sat, 24 Oct 2026 15:25:46 +0000</pubDate>
      <link>https://dev.to/kaushanikoner20/ssorgdev-i-built-a-local-ai-librarian-for-all-those-screenshots-i-keep-forgetting-about-2bhd</link>
      <guid>https://dev.to/kaushanikoner20/ssorgdev-i-built-a-local-ai-librarian-for-all-those-screenshots-i-keep-forgetting-about-2bhd</guid>
      <description>&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;We all have that one folder: hundreds of screenshots, useful information buried somewhere, and absolutely no idea where to find a particular image when we need it.&lt;/p&gt;

&lt;p&gt;A screenshot of a coding error. A tutorial I wanted to try. An important reference. A piece of information I saved because I thought, “I'll need this later.”&lt;/p&gt;

&lt;p&gt;And then I never find it again.&lt;/p&gt;

&lt;p&gt;That's why I built &lt;strong&gt;ssorg.dev&lt;/strong&gt; — a local-first, AI-powered screenshot organizer that helps turn a messy screenshot collection into a searchable personal knowledge library.&lt;/p&gt;

&lt;p&gt;Instead of manually opening images one by one, ssorg.dev helps you find the information inside them.&lt;/p&gt;

&lt;p&gt;Here's what it can do:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Extract text with OCR:&lt;/strong&gt; Reads text from screenshots using Tesseract.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Organize with local AI:&lt;/strong&gt; Uses Gemma 3 4B through Ollama to help generate summaries and structured metadata.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Build a searchable library:&lt;/strong&gt; Stores screenshot records and metadata in SQLite for future searches.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Find screenshots by keywords:&lt;/strong&gt; Search extracted text and browse matching results.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ask ssorg:&lt;/strong&gt; Ask questions about matching screenshots and receive answers grounded in retrieved records, with source references.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Detect exact duplicates:&lt;/strong&gt; Uses SHA-256 hashing to identify identical image files.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The goal isn't to add more work to my workflow. It's to make the information I've already saved easier to find.&lt;/p&gt;

&lt;h2&gt;
  
  
  🎥 Demo
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Watch the ssorg.dev demo video:&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;▶️ &lt;a href="https://github.com/kaushanikoner20-hub/ssorg.dev/blob/main/docs/videos/ssorg-dev-demo.mp4" rel="noopener noreferrer"&gt;Watch the ssorg.dev Demo&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The video demonstrates the project in action, including its screenshot organization and search workflow.&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;GitHub Repository:&lt;/strong&gt; &lt;a href="https://github.com/kaushanikoner20-hub/ssorg.dev" rel="noopener noreferrer"&gt;https://github.com/kaushanikoner20-hub/ssorg.dev&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The project is open source and licensed under the MIT License.&lt;/p&gt;

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

&lt;p&gt;I wanted to build something practical using open-source AI while keeping my screenshots and their contents on my own machine.&lt;/p&gt;

&lt;p&gt;My stack includes:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Python&lt;/strong&gt; — application logic&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Streamlit&lt;/strong&gt; — user interface&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Tesseract OCR and pytesseract&lt;/strong&gt; — text extraction from screenshots&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Gemma 3 4B&lt;/strong&gt; — local AI-assisted classification, summaries, and metadata generation&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ollama&lt;/strong&gt; — running the model locally&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;SQLite&lt;/strong&gt; — persistent storage for screenshot records and metadata&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pillow&lt;/strong&gt; — image handling&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;SHA-256&lt;/strong&gt; — exact duplicate detection&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The workflow is straightforward:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Upload screenshots.&lt;/li&gt;
&lt;li&gt;Extract their text using OCR.&lt;/li&gt;
&lt;li&gt;Use the local Gemma model to help interpret and organize the extracted information.&lt;/li&gt;
&lt;li&gt;Save the results in a local SQLite database and screenshot library.&lt;/li&gt;
&lt;li&gt;Search the collection or ask a question about relevant screenshots.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;For question answering, ssorg.dev retrieves matching records and uses their content as evidence. It also validates source references so answers can be connected to the records they came from.&lt;/p&gt;

&lt;p&gt;I designed the project around local processing rather than a cloud AI API. This makes privacy a core design choice, not an afterthought.&lt;/p&gt;

&lt;p&gt;Building it also meant thinking beyond the happy path: handling processing errors, validating uploads, confirming deletions, detecting duplicates, and avoiding unsupported answers when relevant evidence cannot be found.&lt;/p&gt;

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

&lt;p&gt;Open innovation matters because useful AI tools shouldn't have to depend on expensive services, closed platforms, or sending personal information to external providers.&lt;/p&gt;

&lt;p&gt;With ssorg.dev, I wanted to explore how an open-weight model like Gemma can support a small but meaningful everyday workflow.&lt;/p&gt;

&lt;p&gt;Screenshots can contain code, personal notes, account details, study materials, and other sensitive information. Keeping the AI workflow local gives users more control over where that information is processed.&lt;/p&gt;

&lt;p&gt;The project also brings together accessible technologies that developers can inspect, learn from, and extend.&lt;/p&gt;

&lt;p&gt;For me, open-source AI is not only about making a model available. It's about giving people the ability to build useful tools around it and adapt those tools to their own needs.&lt;/p&gt;

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

&lt;p&gt;The idea behind &lt;em&gt;Slow and Steady&lt;/em&gt; fits this project perfectly.&lt;/p&gt;

&lt;p&gt;I don't want to spend time naming every screenshot, sorting every image into folders, or scrolling endlessly to rediscover something I saved weeks ago.&lt;/p&gt;

&lt;p&gt;I'd rather let the computer handle the repetitive parts and spend my time doing something more useful.&lt;/p&gt;

&lt;p&gt;ssorg.dev is my attempt to make that possible: extract the text, organize the information, and make it searchable when I need it.&lt;/p&gt;

&lt;p&gt;I haven't quantified time saved in a formal benchmark, so I won't claim a specific productivity improvement. The benefit I'm aiming for is simple: less manual searching and a better chance of finding information I've already captured.&lt;/p&gt;

&lt;p&gt;It's a small tool for a familiar problem, built with local AI and open-source technology.&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Best Use of Gemma&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;ssorg.dev uses Gemma 3 4B through Ollama for local AI-assisted screenshot organization and metadata generation. This is the category that best matches the project's actual implementation.&lt;/p&gt;

&lt;h2&gt;
  
  
  Closing Thoughts
&lt;/h2&gt;

&lt;p&gt;This project started with a simple question:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What if saving a screenshot meant I could actually find it again?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Building ssorg.dev has been an opportunity to explore local AI, OCR, retrieval-grounded question answering, and privacy-conscious application design in one practical project.&lt;/p&gt;

&lt;p&gt;I'd love to hear from you:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How do you manage the screenshots, notes, and bits of information you keep saving for later?&lt;/strong&gt; Would you use a local AI librarian for them?&lt;/p&gt;

&lt;p&gt;If you're interested, check out the demo and explore the code on GitHub. Contributions, feedback, and ideas are welcome!&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>productivity</category>
      <category>ai</category>
    </item>
    <item>
      <title>BillBuddy AI: Private Expense Tracking for Freelancers, Powered by Gemma and Ollama</title>
      <dc:creator>Kaushani Koner</dc:creator>
      <pubDate>Sun, 18 Oct 2026 15:56:03 +0000</pubDate>
      <link>https://dev.to/kaushanikoner20/billbuddy-ai-private-expense-tracking-for-freelancers-powered-by-gemma-and-ollama-6gc</link>
      <guid>https://dev.to/kaushanikoner20/billbuddy-ai-private-expense-tracking-for-freelancers-powered-by-gemma-and-ollama-6gc</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a 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;&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Every freelancer knows the pile: photos of receipts, PDF invoices, and a spreadsheet that never gets updated until tax time. I built &lt;strong&gt;BillBuddy AI&lt;/strong&gt; to turn that pile into organized expenses, without uploading a single receipt to someone else's server.&lt;/p&gt;

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

&lt;p&gt;BillBuddy AI is a receipt and invoice organizer that runs entirely on your own computer.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;You drop in a receipt (JPG, PNG, or PDF).&lt;/li&gt;
&lt;li&gt;OCR reads the text.&lt;/li&gt;
&lt;li&gt;An open-weight model, Gemma, suggests the vendor, date, and total.&lt;/li&gt;
&lt;li&gt;You check and correct the fields, then save them.&lt;/li&gt;
&lt;li&gt;The app keeps a searchable expense list with totals per currency, a category breakdown, and a CSV export.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;strong&gt;Who would pay, and why.&lt;/strong&gt; Freelancers and small shop owners spend hours each month typing receipts into spreadsheets. Many of them are also uneasy about sending financial documents to a cloud AI service, because those documents hold client names, addresses, and purchase history. A tool that does the typing locally removes both problems. The side hustle angle I'm exploring is offering receipt and expense organizing to small local businesses: use BillBuddy to cut the data entry, and charge per batch or per month.&lt;/p&gt;

&lt;p&gt;I want to be upfront: this is a working prototype. Nobody has paid for it yet, and I haven't tested whether people would.&lt;/p&gt;

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

&lt;p&gt;BillBuddy runs on your own computer (that is the whole point), so there is no hosted link. Here is a 2-minute walkthrough of the full flow, from reading a receipt to saving, searching, and deleting expenses:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;&lt;a href="https://github.com/kaushanikoner20-hub/BillBuddyAI/blob/main/docs/demo.mp4" rel="noopener noreferrer"&gt;▶ Watch the 2-minute demo&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;[Drag your screenshots into the DEV editor here: the receipt-style OCR text panel, the AI-suggested fields with a warning visible, and the saved expenses with totals in two currencies. Fake receipts only.]&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/kaushanikoner20-hub" rel="noopener noreferrer"&gt;
        kaushanikoner20-hub
      &lt;/a&gt; / &lt;a href="https://github.com/kaushanikoner20-hub/BillBuddyAI" rel="noopener noreferrer"&gt;
        BillBuddyAI
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      
    &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;BillBuddy AI&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;&lt;a rel="noopener noreferrer" href="https://github.com/kaushanikoner20-hub/BillBuddyAI/actions/workflows/ci.yml/badge.svg"&gt;&lt;img src="https://github.com/kaushanikoner20-hub/BillBuddyAI/actions/workflows/ci.yml/badge.svg" alt="CI"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;An open-source, &lt;strong&gt;local-first&lt;/strong&gt; receipt and invoice organizer for freelancers and small businesses
Read a receipt with OCR, let an open-weight AI model (Gemma, running on your own computer through Ollama) suggest the expense details, correct them, and save them to a local SQLite database.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Screenshots&lt;/h2&gt;
&lt;/div&gt;
&lt;p&gt;&lt;a rel="noopener noreferrer" href="https://github.com/kaushanikoner20-hub/BillBuddyAI/docs/screenshots/Screenshot%202026-10-18%20171457.png"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fraw.githubusercontent.com%2Fkaushanikoner20-hub%2FBillBuddyAI%2FHEAD%2Fdocs%2Fscreenshots%2FScreenshot%25202026-10-18%2520171457.png" alt="Screenshot 2026-10-18 171457"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;&lt;a rel="noopener noreferrer" href="https://github.com/kaushanikoner20-hub/BillBuddyAI/docs/screenshots/Screenshot%202026-10-18%20171512.png"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fraw.githubusercontent.com%2Fkaushanikoner20-hub%2FBillBuddyAI%2FHEAD%2Fdocs%2Fscreenshots%2FScreenshot%25202026-10-18%2520171512.png" alt="Screenshot 2026-10-18 171512"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;&lt;a rel="noopener noreferrer" href="https://github.com/kaushanikoner20-hub/BillBuddyAI/docs/screenshots/Screenshot%202026-10-18%20171527.png"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fraw.githubusercontent.com%2Fkaushanikoner20-hub%2FBillBuddyAI%2FHEAD%2Fdocs%2Fscreenshots%2FScreenshot%25202026-10-18%2520171527.png" alt="Screenshot 2026-10-18 171527"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;&lt;a rel="noopener noreferrer" href="https://github.com/kaushanikoner20-hub/BillBuddyAI/docs/screenshots/Screenshot%202026-10-18%20172157.png"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fraw.githubusercontent.com%2Fkaushanikoner20-hub%2FBillBuddyAI%2FHEAD%2Fdocs%2Fscreenshots%2FScreenshot%25202026-10-18%2520172157.png" alt="Screenshot 2026-10-18 172157"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;&lt;a rel="noopener noreferrer" href="https://github.com/kaushanikoner20-hub/BillBuddyAI/docs/screenshots/Screenshot%202026-10-18%20172822.png"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fraw.githubusercontent.com%2Fkaushanikoner20-hub%2FBillBuddyAI%2FHEAD%2Fdocs%2Fscreenshots%2FScreenshot%25202026-10-18%2520172822.png" alt="Screenshot 2026-10-18 172822"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Demo&lt;/h2&gt;

&lt;/div&gt;
&lt;p&gt;&lt;a href="https://github.com/kaushanikoner20-hub/BillBuddyAI/docs/demo.mp4" rel="noopener noreferrer"&gt;Watch the 2-minute demo&lt;/a&gt;&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Features&lt;/h2&gt;

&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;Upload JPG, JPEG, PNG, or PDF receipts (10 MB limit, first 5 PDF pages are read)&lt;/li&gt;
&lt;li&gt;Read the text with Tesseract OCR, including scanned PDFs&lt;/li&gt;
&lt;li&gt;Gemma suggests the vendor, date, and total. The currency is detected from the receipt text itself.&lt;/li&gt;
&lt;li&gt;Warnings when a suggested value does not appear in the receipt text&lt;/li&gt;
&lt;li&gt;Review and correct every field, or fill in the form manually&lt;/li&gt;
&lt;li&gt;Save expenses (vendor, date, total, currency, category, description) in SQLite&lt;/li&gt;
&lt;li&gt;Search by vendor or description, filter by category and date, edit, and delete with confirmation&lt;/li&gt;
&lt;li&gt;Dashboard: number of receipts, totals &lt;strong&gt;per currency&lt;/strong&gt;…&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/kaushanikoner20-hub/BillBuddyAI" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;The repository has Windows setup instructions, a &lt;code&gt;.env.example&lt;/code&gt;, automated tests, a GitHub Actions workflow, and an MIT license.&lt;/p&gt;

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



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Receipt (JPG / PNG / PDF)
   -&amp;gt; Tesseract OCR (PDF pages rendered with PyMuPDF first)
   -&amp;gt; Gemma through Ollama, running locally, suggests fields
   -&amp;gt; validation and warnings
   -&amp;gt; you review and correct
   -&amp;gt; saved in SQLite (reviewed fields only)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Open-weight model:&lt;/strong&gt; Gemma 3 (&lt;code&gt;gemma3:4b&lt;/code&gt;) served locally by &lt;strong&gt;Ollama&lt;/strong&gt;. The backend calls Ollama's local API with temperature 0 and JSON output.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;OCR:&lt;/strong&gt; Tesseract with Pillow. Scanned PDFs work too, because each page is rendered to an image first.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Backend:&lt;/strong&gt; Python and FastAPI. &lt;strong&gt;Storage:&lt;/strong&gt; SQLite, with only the fields you reviewed, not the receipt file or the full OCR text.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Frontend:&lt;/strong&gt; React and Vite.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I built it in stages with help from an AI coding assistant, testing each stage on my own machine before moving on.&lt;/p&gt;

&lt;h3&gt;
  
  
  The lesson: a small model fills in blanks
&lt;/h3&gt;

&lt;p&gt;The part that taught me the most was the currency field. I tested an Indian bill, and the app labeled the amount in &lt;strong&gt;USD&lt;/strong&gt;. Nothing on the bill said dollars. The 4B model simply guessed, because a blank is uncomfortable for a language model. Tesseract's default settings can also misread the ₹ sign, so the model often never saw it.&lt;/p&gt;

&lt;p&gt;So I stopped asking the model about currency. BillBuddy now detects it from the receipt text itself: a currency code, &lt;code&gt;₹&lt;/code&gt;, &lt;code&gt;Rs&lt;/code&gt;, or Indian tax terms like GSTIN and CGST. When the text doesn't say, the field stays empty. A plain &lt;code&gt;$&lt;/code&gt; isn't enough either, because it doesn't tell you the country. An empty field you fill in yourself is better than a confident wrong one.&lt;/p&gt;

&lt;p&gt;That one bug shaped how the whole app treats the model:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Never trust the output.&lt;/strong&gt; Dates must be real dates, amounts must be non-negative numbers, and anything else becomes empty.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Warn when the numbers don't appear in the receipt.&lt;/strong&gt; If the model suggests a total that isn't in the OCR text, the app says so.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Treat receipt text as untrusted.&lt;/strong&gt; A receipt can contain instructions meant to steer the model, so the prompt tells it to ignore them, and the answer is validated anyway.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Never mix currencies.&lt;/strong&gt; Totals are grouped by currency, so rupees and dollars are never added together.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Export safely.&lt;/strong&gt; CSV cells that start like a spreadsheet formula are neutralized.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Because everything is local, testing was easy too: the tests replace the model with a stand-in, so the suite runs without Ollama, and a GitHub Actions workflow runs the tests and the frontend build on every push.&lt;/p&gt;

&lt;h3&gt;
  
  
  Limits
&lt;/h3&gt;

&lt;p&gt;The AI can be wrong, so a person reviews every field before saving. I haven't measured extraction accuracy on real receipts, so I won't claim a number. There is no login or database encryption, so it is meant for one local user, and I developed and tested it on Windows.&lt;/p&gt;

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

&lt;p&gt;A receipt is exactly the kind of document you don't want to hand to a third party: it has names, addresses, and what you bought and where. With an open-weight model running locally, the receipt text never leaves the computer. A closed cloud API can promise to be careful with it, but it still has to receive it.&lt;/p&gt;

&lt;p&gt;Open also changed the economics. There are no API keys and no per-receipt fee, so it costs nothing extra whether a shop processes ten receipts or ten thousand. For a business that might run this every day, that is the difference between a tool that works and one that eats the margin. After the one-time downloads, the pipeline doesn't need the internet.&lt;/p&gt;

&lt;p&gt;And the model is one setting. If a larger Gemma, or a different model, reads receipts better, I change a single line instead of rewriting the app or being locked into someone's pricing.&lt;/p&gt;

&lt;p&gt;I'll be honest about the trade-off: a 4B model on a laptop is less accurate than a large hosted model, and that is why the review step and the validation exist. For this use case, I think a reviewed, private, free-to-run answer beats a faster one that needs your documents uploaded.&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; BillBuddy runs Gemma 3 (4B) locally through Ollama to read OCR text and suggest expense fields, with validation and warnings built around its answers.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>opensource</category>
      <category>productivity</category>
    </item>
    <item>
      <title>TrailBuddy: The AI That Disappears—Powered by Gemma, Built to Get You Outside</title>
      <dc:creator>Kaushani Koner</dc:creator>
      <pubDate>Fri, 09 Oct 2026 19:21:39 +0000</pubDate>
      <link>https://dev.to/kaushanikoner20/trailbuddy-the-ai-that-disappears-powered-by-gemma-built-to-get-you-outside-27j3</link>
      <guid>https://dev.to/kaushanikoner20/trailbuddy-the-ai-that-disappears-powered-by-gemma-built-to-get-you-outside-27j3</guid>
      <description>&lt;h2&gt;
  
  
  TrailBuddy: The AI That Disappears
&lt;/h2&gt;

&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week1-2026-10-05"&gt;Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;TrailBuddy turns a few choices—your mood, available time, activity, and preferred difficulty—into a personalized outdoor adventure.&lt;/p&gt;

&lt;p&gt;It isn’t meant to be an AI companion you keep checking while you’re outside. Gemma designs a short sequence of sensory missions: listen for a sound you usually miss, notice a pattern, or explore your surroundings with curiosity. TrailBuddy then asks whether you’re going out, shows a closing scene where a bird takes off, and invites you to put your phone away.&lt;/p&gt;

&lt;p&gt;The idea behind it is simple:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;“What if the best thing an AI could do for you was make you stop looking at it?”&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;After generating and preparing an adventure, you can continue through its missions offline. Progress and an optional reflection are saved in your browser. Generating a &lt;em&gt;new&lt;/em&gt; adventure still requires the configured backend and Ollama/Gemma to be available.&lt;/p&gt;

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

&lt;p&gt;&lt;a href="https://github.com/kaushanikoner20-hub/TrailBuddy/blob/codex/stage-4-offline-first/assests/Screen%20Recording%202026-10-09%20204549.mp4" rel="noopener noreferrer"&gt;Watch the TrailBuddy screen recording&lt;/a&gt;&lt;/p&gt;

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

&lt;p&gt;&lt;a href="https://github.com/kaushanikoner20-hub/TrailBuddy" rel="noopener noreferrer"&gt;TrailBuddy on GitHub&lt;/a&gt;&lt;/p&gt;

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

&lt;p&gt;TrailBuddy uses &lt;strong&gt;Gemma 3 (&lt;code&gt;gemma3:4b&lt;/code&gt;)&lt;/strong&gt; through &lt;strong&gt;Ollama&lt;/strong&gt; for local adventure generation. The React and Vite frontend sends the selected preferences to a Node.js and Express backend, which requests a structured adventure from the model and validates the response before showing it.&lt;/p&gt;

&lt;p&gt;The frontend presents the generated missions one at a time in Phone Away Mode. The closing scene uses Three.js, and the production app includes a service worker that caches the application shell and assets for a prepared offline adventure. Local progress and reflections are stored in browser storage.&lt;/p&gt;

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

&lt;p&gt;The open-weight Gemma model and local Ollama setup let me build around a model I can run and experiment with on my own machine. I could shape the prompt for a specific purpose—an outdoor experience designer—and validate the result as structured missions instead of building another general-purpose chat interface.&lt;/p&gt;

&lt;p&gt;Local inference also means the default setup doesn’t need a cloud LLM API to generate an adventure. The offline flow goes further: once an adventure is prepared, the user can leave the AI and backend behind and continue with the saved missions. That distinction matters: a &lt;em&gt;new&lt;/em&gt; adventure still needs the configured model and backend.&lt;/p&gt;

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


&lt;div class="ltag-agent-session"&gt;
  &lt;div class="agent-session-header"&gt;
    
    &lt;span class="agent-session-tool-icon-badge" title="Codex"&gt;
&lt;/span&gt;
    &lt;span class="agent-session-title"&gt;TrailBuddy Hacktoberfest Submission&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;Curate and save an Agent Session for my TrailBuddy Hacktoberfest submission. Remove secrets, credentials, and private machine paths. Save the session only—don’t create or publish my DEV article. Return the session ID or slug and the exact {% agent_session ... %} tag.&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;Saved this curated Agent Session for the TrailBuddy Hacktoberfest submission. No secrets, credentials, or private machine paths are included.&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;
        2 of 2 messages
    &lt;/span&gt;
  &lt;/div&gt;
&lt;/div&gt;


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

&lt;ul&gt;
&lt;li&gt;Best Use of Gemma&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>opensource</category>
      <category>showdev</category>
    </item>
    <item>
      <title>New to community!</title>
      <dc:creator>Kaushani Koner</dc:creator>
      <pubDate>Fri, 02 Oct 2026 19:30:03 +0000</pubDate>
      <link>https://dev.to/kaushanikoner20/new-to-community-1fjp</link>
      <guid>https://dev.to/kaushanikoner20/new-to-community-1fjp</guid>
      <description>&lt;p&gt;hello everyone&lt;/p&gt;

</description>
    </item>
  </channel>
</rss>
