HF chaallenges

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady

What I Built

Demo

Code

How I Built It

Why Does Open Innovation Matter?

My Agent Session

The Lazy Payoff

Prize Categories

You are a senior Python engineer helping me build a hackathon project in 1-2 days.

Project: LazyLecture

A local, offline tool that turns lecture files into study material automatically.
I drop a PDF (or audio file) into an "inbox" folder. A watcher detects it, a local
open-weight LLM processes it, and the output appears in an "output" folder:

1. summary.md - a one-page summary with headings
2. flashcards.md and flashcards.csv (Anki-importable: front,back)
3. quiz.md - 5 multiple-choice questions with answers and one-line explanations

Hard constraints

• Everything runs locally and offline. No paid or cloud APIs.
• LLM via Ollama (default model: qwen2.5:7b; model name must be a config setting so I can swap in llama3.1:8b or a 3B model and compare).
• Python 3.11+, libraries: watchdog, pypdf, requests (or the ollama python client), pyyaml. Optional: whisper.cpp for audio transcription.
• Must work on Windows, macOS and Linux (use pathlib, no shell-specific code).
• Keep it simple: no frameworks beyond what's listed, minimal files, readable code.

Architecture

• config.yaml: model name, inbox/output paths, chunk size, number of flashcards/questions
• extract.py: PDF -> clean text (handle empty or scanned pages gracefully)
• chunk.py: split text into chunks that fit the model's context window
• llm.py: wrapper around Ollama with retries and timeouts
• pipeline.py: chunk -> summarize each chunk -> merge into final summary -> generate flashcards -> generate quiz. Ask the model for JSON for flashcards/quiz, validate it, and retry once if the JSON is malformed.
• watcher.py: watchdog loop that processes new files, waits until the file has finished copying, and skips files already processed.
• main.py: CLI entry point (python main.py watch and python main.py run <file>)
• Optional stretch: app.py, a small Streamlit page to take the quiz and track wrong answers.

How to work

1. First, show me the plan and file tree in under 15 lines. Wait for my OK.
2. Build in small steps: (a) extract + chunk, (b) LLM wrapper, (c) pipeline, (d) watcher, (e) outputs. After each step, tell me exactly how to run and test it.
3. Write a few pytest tests for chunking and JSON validation, with the LLM mocked.
4. Log clearly (which file, which stage, time taken per stage). I want timing data for my write-up.
5. Add a benchmark script that runs the same PDF through two different models and prints time taken and output length, so I can compare speed vs. quality.
6. Finish with a README: what it does, setup steps (install Ollama, pull model, pip install), usage, a "why open-source/local matters" section, and troubleshooting.

Quality bar

• A new user can go from clone to working output in under 10 minutes.
• Errors never crash the watcher; they are logged and the file is moved to a "failed" folder.
• Explain any non-obvious decision in one sentence. Don't add features I didn't ask for.

Start with step 1 (the plan).