<?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: Rajdeep Dolui</title>
    <description>The latest articles on DEV Community by Rajdeep Dolui (@students_room_ef2650e0bc).</description>
    <link>https://dev.to/students_room_ef2650e0bc</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%2F4152732%2Fdc7baea1-5028-49fd-8ba8-c6e9518bcbe6.jpg</url>
      <title>DEV Community: Rajdeep Dolui</title>
      <link>https://dev.to/students_room_ef2650e0bc</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/students_room_ef2650e0bc"/>
    <language>en</language>
    <item>
      <title>OpenLoom: Watching AI Coding Agents So You Don't Have To</title>
      <dc:creator>Rajdeep Dolui</dc:creator>
      <pubDate>Sun, 01 Nov 2026 09:55:20 +0000</pubDate>
      <link>https://dev.to/students_room_ef2650e0bc/openloom-watching-ai-coding-agents-so-you-dont-have-to-21cf</link>
      <guid>https://dev.to/students_room_ef2650e0bc/openloom-watching-ai-coding-agents-so-you-dont-have-to-21cf</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;AI coding agents are great until one gets stuck in a loop and you only notice an hour later. OpenLoom is an open-source, webhook-driven task harness and observer for OpenCode that fixes this.&lt;/p&gt;

&lt;p&gt;You send a goal to a simple task API. OpenLoom attaches it to an OpenCode session, watches the session for activity, and streams updates as they happen. It doesn't replace OpenCode. It fills the gaps in OpenCode's HTTP API with:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;a JSON task API&lt;/li&gt;
&lt;li&gt;a session monitor&lt;/li&gt;
&lt;li&gt;webhook notifications&lt;/li&gt;
&lt;li&gt;a web dashboard showing live tasks, per-session token usage, and pending tool approvals (Once / Always / Deny)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If an agent gets stuck, you can abort its session and send a fresh prompt, even from your phone.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;My part:&lt;/strong&gt; I added a new feature to OpenLoom, working solo inside an existing open-source codebase. I built it for the &lt;strong&gt;Best Open-Source AI Project&lt;/strong&gt; challenge in the Hacktoberfest hackathon, submitted through MLH.&lt;/p&gt;

&lt;p&gt;Repo: &lt;a href="https://github.com/bidyutdolui1970-cpu/my-public-repo" rel="noopener noreferrer"&gt;https://github.com/bidyutdolui1970-cpu/my-public-repo&lt;/a&gt;&lt;/p&gt;

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

&lt;p&gt;OpenLoom talks to OpenCode over HTTP. The stack is Python with a FastAPI backend and a Svelte dashboard. I didn't use any MLH partner tools, so the main technologies were OpenCode and the open-source Python tooling around it.&lt;/p&gt;

&lt;p&gt;Getting started takes three commands:&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;pip install "openloom[ui]"
opencode serve
openloom serve
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;p&gt;Then open the dashboard at &lt;a href="http://127.0.0.1:55413" rel="noopener noreferrer"&gt;http://127.0.0.1:55413&lt;/a&gt;.&lt;/p&gt;

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

&lt;p&gt;I built this solo during Hacktoberfest 2026, entering the Best Open-Source AI Project challenge through MLH. Working alone meant I owned every step: reading the codebase, designing the feature, and testing it.&lt;/p&gt;

&lt;p&gt;The hardest part was debugging. OpenLoom sits between your tools and an AI agent, so when something broke it was hard to tell which side the fault was on. Tracing it step by step taught me to check what the system is actually doing instead of trusting what I assumed it was doing.&lt;/p&gt;

&lt;p&gt;What I'll remember most: contributing to a real open-source project is mostly careful reading and patient debugging, and the feature itself is the smaller part.&lt;/p&gt;

</description>
      <category>mlhacks</category>
      <category>devchallenge</category>
      <category>hackathon</category>
    </item>
    <item>
      <title>Itihaser Bhoot: An Offline Ghost of History</title>
      <dc:creator>Rajdeep Dolui</dc:creator>
      <pubDate>Tue, 27 Oct 2026 03:40:35 +0000</pubDate>
      <link>https://dev.to/students_room_ef2650e0bc/itihaser-bhoot-an-offline-ghost-of-history-3o1d</link>
      <guid>https://dev.to/students_room_ef2650e0bc/itihaser-bhoot-an-offline-ghost-of-history-3o1d</guid>
      <description>&lt;p&gt;This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine.&lt;/p&gt;

&lt;p&gt;I am a B.A. History student at Burdwan University. For this challenge, I wanted to build something that connects my interest in History with my interest in AI.&lt;/p&gt;

&lt;p&gt;My project is called Itihaser Bhoot (ইতিহাসের ভূত). It is an offline AI assistant that acts like a ghost of a historical figure. It answers History questions using only my own notes.&lt;/p&gt;

&lt;p&gt;The idea is simple: an AI that knows what I have studied, shows where its answers come from, and admits when it does not know something.&lt;/p&gt;

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

&lt;p&gt;Itihaser Bhoot is a local History question-answering system.&lt;/p&gt;

&lt;p&gt;It has three main goals:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Answer questions using only my personal History notes.&lt;/li&gt;
&lt;li&gt;Quote the source passage so I can check the answer.&lt;/li&gt;
&lt;li&gt;Stay silent or refuse to answer when the notes do not contain enough information.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I wanted to avoid answers that sound convincing but are not supported by my study material.&lt;/p&gt;

&lt;p&gt;Demo&lt;/p&gt;

&lt;p&gt;The project runs locally on my laptop.&lt;/p&gt;

&lt;p&gt;Example workflow:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;I ask a History question.&lt;/li&gt;
&lt;li&gt;The system searches my notes for relevant passages.&lt;/li&gt;
&lt;li&gt;Gemma generates an answer based on the retrieved material.&lt;/li&gt;
&lt;li&gt;The answer includes the source passage.&lt;/li&gt;
&lt;li&gt;If the notes do not cover the question, the system refuses to answer.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The goal is not to make the AI sound knowledgeable about everything. It is to make it reliable within the limits of my notes.&lt;/p&gt;

&lt;p&gt;Code&lt;/p&gt;

&lt;p&gt;GitHub repository: &lt;a href="https://github.com/bidyutdolui1970-cpu/itihaser-bhoot" rel="noopener noreferrer"&gt;https://github.com/bidyutdolui1970-cpu/itihaser-bhoot&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The repository contains the code and instructions for running the project locally.&lt;/p&gt;

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

&lt;p&gt;I used Gemma 3 1B through Ollama for local language generation.&lt;/p&gt;

&lt;p&gt;For retrieving relevant passages from my notes, I used BM25. It helps find text that matches the question before the model generates an answer.&lt;/p&gt;

&lt;p&gt;The system follows a retrieval-based approach:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;My notes are the knowledge source.&lt;/li&gt;
&lt;li&gt;BM25 retrieves relevant passages.&lt;/li&gt;
&lt;li&gt;Gemma 3 1B generates an answer from the retrieved context.&lt;/li&gt;
&lt;li&gt;The source passage is included for verification.&lt;/li&gt;
&lt;li&gt;Questions without sufficient supporting material should be refused.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I built this on a laptop with 3.6 GB of RAM.&lt;/p&gt;

&lt;p&gt;I wanted to see whether a small local model could be useful for a focused academic task without relying on a cloud API.&lt;/p&gt;

&lt;p&gt;Why Open Innovation Matters&lt;/p&gt;

&lt;p&gt;For me, open innovation means making useful technology accessible to people who do not have powerful computers or paid AI subscriptions.&lt;/p&gt;

&lt;p&gt;Itihaser Bhoot has a few practical advantages:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Offline: It runs without an internet connection once the required model is available locally.&lt;/li&gt;
&lt;li&gt;Free per question: There is no cloud API charge for asking questions.&lt;/li&gt;
&lt;li&gt;Runs on my laptop: I do not need expensive hardware for this focused task.&lt;/li&gt;
&lt;li&gt;Private notes: My study notes stay on my own device.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;As a student, these things matter to me. I want to experiment with AI while keeping control over my own study material.&lt;/p&gt;

&lt;p&gt;What Went Wrong&lt;/p&gt;

&lt;p&gt;The project is still experimental, and I faced several problems.&lt;/p&gt;

&lt;p&gt;The biggest challenge was answer quality. A small model can sometimes misunderstand a retrieved passage or produce an answer that goes beyond the available evidence.&lt;/p&gt;

&lt;p&gt;Bengali also introduced additional difficulties. Bengali questions and notes do not always produce reliable retrieval results. The model can struggle with Bengali phrasing, historical terminology, and generating natural Bengali answers.&lt;/p&gt;

&lt;p&gt;Another challenge is refusal behavior. Even when the notes do not contain an answer, the model may try to generate one. I need to improve the system's ability to recognize when the retrieved evidence is insufficient.&lt;/p&gt;

&lt;p&gt;I am still testing these limitations rather than claiming they are solved.&lt;/p&gt;

&lt;p&gt;Prize Category: Best Use of Gemma&lt;/p&gt;

&lt;p&gt;I am submitting Itihaser Bhoot for Best Use of Gemma.&lt;/p&gt;

&lt;p&gt;I used Gemma 3 1B as a local language model for a specific academic purpose. The project is not intended to replace textbooks, teachers, or historical research.&lt;/p&gt;

&lt;p&gt;It is an experiment in building a small, private, and evidence-based History assistant using open tools.&lt;/p&gt;

&lt;p&gt;I hope to improve its retrieval quality, Bengali support, and refusal behavior as I continue working on it.&lt;/p&gt;

&lt;h1&gt;
  
  
  Hacktoberfest #devchallenge #hf26challenge #opensource #ai
&lt;/h1&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>Dhire Dhire: a slow, offline History study helper for a 3.6 GB laptop tags: devchallenge, hf26challenge</title>
      <dc:creator>Rajdeep Dolui</dc:creator>
      <pubDate>Tue, 20 Oct 2026 06:34:51 +0000</pubDate>
      <link>https://dev.to/students_room_ef2650e0bc/dhire-dhire-a-slow-offline-history-study-helper-for-a-36-gb-laptop-tags-devchallenge-d34</link>
      <guid>https://dev.to/students_room_ef2650e0bc/dhire-dhire-a-slow-offline-history-study-helper-for-a-36-gb-laptop-tags-devchallenge-d34</guid>
      <description>&lt;p&gt;This is a submission for the Hacktoberfest Open-Source AI Challenge: Week 3 — Slow and Steady&lt;br&gt;
What I Built&lt;br&gt;
I'm a History student at Burdwan University. My laptop has about 3.6 GB of RAM, so most AI tools either need the internet or need more memory than I have.&lt;br&gt;
For "Slow and Steady," I built around that limit instead of fighting it. Dhire Dhire (ধীরে ধীরে, "slowly, slowly") is an offline study helper. You give it your notes or textbook text. It cuts the text into small chunks and works through them one at a time with Gemma 3 1B, turning each chunk into exam-style questions with short answers.&lt;br&gt;
After every chunk it saves a checkpoint. If my laptop sleeps or crashes, I run it again and it continues from where it stopped. I can start it at night and read the results in the morning.&lt;br&gt;
It's for students like me: limited hardware, patchy internet, a lot of reading.&lt;br&gt;
Code&lt;br&gt;
&lt;a href="https://github.com/bidyutdolui1970-cpu/dhire-dhire" rel="noopener noreferrer"&gt;https://github.com/bidyutdolui1970-cpu/dhire-dhire&lt;/a&gt;&lt;br&gt;
It's plain Python with no dependencies, plus Ollama for running the model. I built it during the Week 3 window, with AI assistance for writing the code and tests.&lt;br&gt;
How I Built It&lt;br&gt;
Model: Gemma 3 1B (gemma3:1b), run locally through Ollama with a small context window to keep memory low.&lt;br&gt;
Chunking: text is split at sentence ends, including the Bengali danda (।), into chunks of about 180 words.&lt;br&gt;
Checkpoints: every finished chunk is saved to SQLite immediately. A chunk only counts as done after its answer is saved, so an interrupted chunk is simply redone. I wrote tests that interrupt a run halfway and check that the restart only does the remaining work.&lt;br&gt;
Output: a Markdown notes file and a CSV of flashcards. If the model ignores the Q/A format, its raw text is kept instead of thrown away.&lt;br&gt;
Measured on my laptop: RAM in use while running: 26 GB. Seconds per chunk: 5 s. Total run time for my test notes:6 min.&lt;br&gt;
Why Open Innovation Matters Here&lt;br&gt;
It runs offline. After the one-time model download, no internet is needed.&lt;br&gt;
It's free. No API bill and no per-token cost, which matters for a student.&lt;br&gt;
It's private. My notes never leave the laptop.&lt;br&gt;
I can swap models. If a better small model appears, I change one setting (Gemma model) and the rest of the pipeline stays the same.&lt;br&gt;
Where it beat a cloud approach: a cloud API would be faster per request, but it needs internet, costs money and sends my material to someone else's server. Slow but always available fits my situation better.&lt;br&gt;
Prize Categories&lt;br&gt;
Best Use of Gemma: Gemma 3 1B is the core of the project and runs fully locally on a 3.6 GB laptop.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>"Dokan Khata: A Voice Notebook for Shopkeepers Who Sell on Credit, Powered by Gemma and Whisper"</title>
      <dc:creator>Rajdeep Dolui</dc:creator>
      <pubDate>Tue, 13 Oct 2026 10:52:52 +0000</pubDate>
      <link>https://dev.to/students_room_ef2650e0bc/dokan-khata-a-voice-notebook-for-shopkeepers-who-sell-on-credit-powered-by-gemma-and-whisper-1mc6</link>
      <guid>https://dev.to/students_room_ef2650e0bc/dokan-khata-a-voice-notebook-for-shopkeepers-who-sell-on-credit-powered-by-gemma-and-whisper-1mc6</guid>
      <description>&lt;p&gt;ROLE&lt;br&gt;
You are an editor helping me write a DEV Community challenge post. The write-up is the most heavily weighted judging criterion, so clarity, honesty and a human voice matter more than polish.&lt;/p&gt;

&lt;p&gt;TASK&lt;br&gt;
Write my submission post for the "Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle". Use exactly this structure:&lt;/p&gt;




&lt;p&gt;title: "Dokan Khata: A Voice Notebook for Shopkeepers Who Sell on Credit, Powered by Gemma and Whisper"&lt;br&gt;
published: false&lt;/p&gt;

&lt;h2&gt;
  
  
  tags: devchallenge, hf26challenge
&lt;/h2&gt;

&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;Then these headings, in order:&lt;/p&gt;

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

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

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

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

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

&lt;h2&gt;
  
  
  My First User (include ONLY if I give real facts below, otherwise omit)
&lt;/h2&gt;

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

&lt;p&gt;FACTS (use only these; never invent anything)&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Project: Dokan Khata, an offline Bengali voice ledger for corner shops. The shopkeeper speaks a credit sale, the app transcribes it, extracts customer/items/amount/due date, asks for confirmation, and updates the customer's balance. It also makes a weekly "who owes me" summary and polite Bengali reminder drafts.&lt;/li&gt;
&lt;li&gt;Problem and who pays: [your words about the shopkeeper problem]&lt;/li&gt;
&lt;li&gt;Models/tools used: faster-whisper small (int8), Gemma 3 1B (4-bit) via Ollama, SQLite, Flask [edit to match what you really used]&lt;/li&gt;
&lt;li&gt;Hardware: low-end laptop, about 3.6 GB RAM, CPU only. Peak RAM measured: [number]&lt;/li&gt;
&lt;li&gt;Measured results: [accuracy per field from my test set, latency per step]&lt;/li&gt;
&lt;li&gt;Things that broke and how I fixed them: [paste from FAILURES.md]&lt;/li&gt;
&lt;li&gt;Demo link/video: [link]&lt;/li&gt;
&lt;li&gt;Repo link: [link]&lt;/li&gt;
&lt;li&gt;First user (delete if none): [who, what they did, what they said, did they say they would pay]&lt;/li&gt;
&lt;li&gt;Categories entered: Best Use of Gemma [add others only if truly used]&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;STYLE&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;First person, plain English, short paragraphs, no hype or buzzword stacking.&lt;/li&gt;
&lt;li&gt;Open with a concrete scene (a shopkeeper with a paper notebook), not a definition.&lt;/li&gt;
&lt;li&gt;In "Why Does Open Innovation Matter?" give specific reasons tied to my facts: customer data stays on the laptop, works offline, zero per-use cost, adaptable to Bengali. Say where a closed API would have been worse.&lt;/li&gt;
&lt;li&gt;Include at least one honest failure and what I learned.&lt;/li&gt;
&lt;li&gt;If a fact is missing, leave a [PLACEHOLDER] instead of guessing.&lt;/li&gt;
&lt;li&gt;Under 900 words.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;AFTER THE DRAFT&lt;br&gt;
List the placeholders I still need to fill, and flag any claim that is weaker than it sounds.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>Fridge Forecast: a weekly outdoor plan you print once and never look up again</title>
      <dc:creator>Rajdeep Dolui</dc:creator>
      <pubDate>Tue, 06 Oct 2026 03:42:41 +0000</pubDate>
      <link>https://dev.to/students_room_ef2650e0bc/fridge-forecast-a-weekly-outdoor-plan-you-print-once-and-never-look-up-again-3apn</link>
      <guid>https://dev.to/students_room_ef2650e0bc/fridge-forecast-a-weekly-outdoor-plan-you-print-once-and-never-look-up-again-3apn</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;Apps that tell you to go outside have an obvious flaw: you have to open an app.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Fridge Forecast&lt;/strong&gt; moves the whole thing onto paper. Once a week, you export your calendar and run one command. It finds the free windows in your week, checks the forecast for each, and prints a one-page card for the fridge showing the best times to go outside and what to do with them. After that, the screen's job is finished.&lt;/p&gt;

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

&lt;p&gt;The flow has three steps:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Export&lt;/strong&gt; your calendar as an &lt;code&gt;.ics&lt;/code&gt; file and add a line about what you like.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Run&lt;/strong&gt; Fridge Forecast. It picks the best windows and writes a suggestion for each.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Print&lt;/strong&gt; the card and put it on the fridge.&lt;/li&gt;
&lt;/ol&gt;

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

&lt;p&gt;Everything runs locally except the forecast lookup:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Calendar:&lt;/strong&gt; the &lt;code&gt;icalendar&lt;/code&gt; library reads your &lt;code&gt;.ics&lt;/code&gt; file and works out free windows.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Forecast:&lt;/strong&gt; Open-Meteo, a free open API that needs no key.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Choosing windows:&lt;/strong&gt; plain code scores each window on rain, wind, temperature, and daylight. The model never decides when you go out.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Suggestions:&lt;/strong&gt; an open-weight Gemma model served through Ollama writes one friendly suggestion per chosen window, using your note about what you like.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Output:&lt;/strong&gt; a printable HTML page.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The key design choice is the split. Code makes the decisions that must be right, such as when the weather is good and when you're free. The model only writes the human part. Its output is validated, and if it fails, fixed fallback text keeps the card complete.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Open Source AI
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Privacy:&lt;/strong&gt; your calendar shows when you're home, where you are, and who you meet. With a local model, it never leaves your machine.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cost:&lt;/strong&gt; it costs nothing to run, so it can be a weekly habit with no subscription.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Control:&lt;/strong&gt; the scoring rules and the prompt are plain files you can edit to suit your own life.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>gemma</category>
      <category>ai</category>
    </item>
    <item>
      <title>OpenLoom: answers from your own documents, with a citation for every claim</title>
      <dc:creator>Rajdeep Dolui</dc:creator>
      <pubDate>Sun, 04 Oct 2026 16:39:24 +0000</pubDate>
      <link>https://dev.to/students_room_ef2650e0bc/openloom-answers-from-your-own-documents-with-a-citation-for-every-claim-4map</link>
      <guid>https://dev.to/students_room_ef2650e0bc/openloom-answers-from-your-own-documents-with-a-citation-for-every-claim-4map</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;Asking an AI about your own documents has an awkward trust problem. It gives a confident answer, and you have no idea whether it came from your files or from the model's imagination.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;OpenLoom&lt;/strong&gt; tackles that. You upload your PDFs and notes, ask a question, and OpenLoom answers using only those documents. Every claim points back to the exact passage it came from, like a thread you can pull. If the answer isn't in your files, it says so instead of guessing.&lt;/p&gt;

&lt;p&gt;The name comes from weaving: many separate threads of source material become one answer, and you can always trace it back.&lt;/p&gt;

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

&lt;p&gt;The flow is simple:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Upload&lt;/strong&gt; one or more PDFs or text files.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ask&lt;/strong&gt; a question in plain language.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Read&lt;/strong&gt; the answer, then expand any citation to see the original passage and the file and page it came from.&lt;/li&gt;
&lt;/ol&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/bidyutdolui1970-cpu" rel="noopener noreferrer"&gt;
        bidyutdolui1970-cpu
      &lt;/a&gt; / &lt;a href="https://github.com/bidyutdolui1970-cpu/my-public-repo" rel="noopener noreferrer"&gt;
        my-public-repo
      &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;OpenLoom&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;OpenLoom is a lightweight platform for building and running AI-powered applications that leverage open models. It helps teams connect model workflows, orchestration logic, and deployment pipelines without locking themselves into a single proprietary provider.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;What OpenLoom does&lt;/h2&gt;
&lt;/div&gt;
&lt;p&gt;OpenLoom gives developers a practical way to:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;orchestrate AI workflows and tool usage&lt;/li&gt;
&lt;li&gt;connect applications to open-source large language models&lt;/li&gt;
&lt;li&gt;manage prompts, data flows, and model calls in a reusable way&lt;/li&gt;
&lt;li&gt;integrate AI features into products while keeping control over model choice and deployment&lt;/li&gt;
&lt;li&gt;support experimentation and iteration across models and environments&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;In simple terms, OpenLoom helps turn model capabilities into usable application features with a clear and flexible architecture.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Where the open model is used&lt;/h2&gt;
&lt;/div&gt;
&lt;p&gt;The open model is used in the model layer of OpenLoom, where it powers inference, reasoning, and task execution in a transparent and portable way. This is typically the place where:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;chat and agent workflows run&lt;/li&gt;
&lt;li&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/bidyutdolui1970-cpu/my-public-repo" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;h2&gt;
  
  
  Open-Source Technologies
&lt;/h2&gt;

&lt;p&gt;OpenLoom is built entirely from open pieces, and all of it runs locally:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Embeddings:&lt;/strong&gt; &lt;code&gt;sentence-transformers&lt;/code&gt; turns document chunks into vectors.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Vector search:&lt;/strong&gt; Chroma stores the chunks and finds the most relevant ones for a question.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Language model:&lt;/strong&gt; an open-weight Gemma model, served through Ollama, writes the answer from the retrieved passages only.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Interface:&lt;/strong&gt; Streamlit for upload, questions, and source viewing.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The key design choice is the prompt. The model is told to answer only from the retrieved passages, cite them by number, and say "not found in your documents" when the evidence isn't there. That one rule is what makes the citations trustworthy.&lt;/p&gt;

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

&lt;p&gt;Retrieval quality matters more than model size. Chunk size and overlap changed answer quality more than anything else, because a chunk that cuts a thought in half gives the model nothing useful to cite. Keeping the file name and page number attached to every chunk made citations nearly free to build.&lt;/p&gt;

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

&lt;p&gt;I built OpenLoom for Hacktoberfest 2026, MLH's open-source AI hackathon, where the goal was to build something new with open-source AI at its core. Building against that constraint pushed me toward local models, and that turned out to be the right call for a tool that handles people's private documents.&lt;/p&gt;

</description>
      <category>mlhacks</category>
      <category>devchallenge</category>
      <category>hackathon</category>
    </item>
    <item>
      <title>Recipe Keeper: turning a family member's voice notes into a recipe book, fully offline</title>
      <dc:creator>Rajdeep Dolui</dc:creator>
      <pubDate>Sun, 04 Oct 2026 16:34:43 +0000</pubDate>
      <link>https://dev.to/students_room_ef2650e0bc/recipe-keeper-turning-a-family-members-voice-notes-into-a-recipe-book-fully-offline-3mo3</link>
      <guid>https://dev.to/students_room_ef2650e0bc/recipe-keeper-turning-a-family-members-voice-notes-into-a-recipe-book-fully-offline-3mo3</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-weekend-2026-10-01"&gt;Hacktoberfest Weekend Challenge: Build for a Friend&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;Some of the best cooks in our families never write anything down. The recipes live in their heads, and the only trace is a rambling voice note that says things like "add salt until it feels right." When that person is gone, or just busy, the recipe goes with them.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Recipe Keeper&lt;/strong&gt; is a small tool I built for a family member who cooks entirely from memory. You upload a voice note of them talking through a dish. The app transcribes it, then a local language model turns the transcript into a clean recipe card with ingredients, steps, and tips. Saved cards are collected into a printable family recipe book.&lt;/p&gt;

&lt;p&gt;The goal was simple: let them keep cooking the way they always have, and let the rest of us finally hold on to the food.&lt;/p&gt;

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

&lt;p&gt;The flow has three steps:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Upload&lt;/strong&gt; a voice note, for example someone explaining how they make a dish.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Review&lt;/strong&gt; the transcript and the generated recipe card side by side, and fix anything the model got wrong.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Save&lt;/strong&gt; the card. All saved recipes are combined into one recipe book you can print.&lt;/li&gt;
&lt;/ol&gt;

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

&lt;p&gt;Everything runs on a single laptop, with no cloud services and no API keys:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Speech-to-text:&lt;/strong&gt; Whisper, run through &lt;code&gt;faster-whisper&lt;/code&gt;. It supports many languages, which matters because spoken recipes often mix languages.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Structuring:&lt;/strong&gt; an open-weight Gemma model served locally with Ollama. I prompt it to return strict JSON with a title, ingredients, steps, and tips, then validate the JSON before showing it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Interface:&lt;/strong&gt; a simple Streamlit app for upload, review, and editing.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Output:&lt;/strong&gt; an HTML recipe book generated from the saved recipes, which can be printed to PDF.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The trickiest design problem is that spoken recipes are vague ("a handful", "until it smells right"). The prompt tells the model to keep the speaker's own wording and never invent exact measurements, and the review step lets a human correct anything before it is saved.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Privacy:&lt;/strong&gt; these are family voice recordings. With local models they never leave the laptop, and no company's server ever hears them.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cost:&lt;/strong&gt; it costs nothing to run, so there is no API bill for something a family might use for years.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Control:&lt;/strong&gt; if transcription is weak for a particular accent or language, I can switch to a larger Whisper model or adjust the prompt. With a closed API, I would be stuck with whatever it gives me.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Offline:&lt;/strong&gt; it works in the kitchen, with no internet connection needed.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;For something as personal as a family member's voice and recipes, keeping everything local made more sense than any cloud service could.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>weekendchallenge</category>
      <category>hf26challenge</category>
    </item>
  </channel>
</rss>
