Dhire Dhire: a slow, offline History study helper for a 3.6 GB laptop tags: devchallenge, hf26challenge

This is a submission for the Hacktoberfest Open-Source AI Challenge: Week 3 — Slow and Steady
What I Built
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.
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.
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.
It's for students like me: limited hardware, patchy internet, a lot of reading.
Code
https://github.com/bidyutdolui1970-cpu/dhire-dhire
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.
How I Built It
Model: Gemma 3 1B (gemma3:1b), run locally through Ollama with a small context window to keep memory low.
Chunking: text is split at sentence ends, including the Bengali danda (।), into chunks of about 180 words.
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.
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.
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.
Why Open Innovation Matters Here
It runs offline. After the one-time model download, no internet is needed.
It's free. No API bill and no per-token cost, which matters for a student.
It's private. My notes never leave the laptop.
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.
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.
Prize Categories
Best Use of Gemma: Gemma 3 1B is the core of the project and runs fully locally on a 3.6 GB laptop.

Comments

[Rajdeep Dolui]

Excited for "Slow and Steady"! I'm building Dhire Dhire, an offline History study helper that runs Gemma 3 1B one chunk at a time on a 3.6 GB RAM laptop. It saves a checkpoint after each chunk, so it can resume after a crash. Slow on purpose, free and private. Repo: github.com/bidyutdolui1970-cpu/dhi...