LectureLens (Mozhi): Ask Your Lectures Anything, Fully Offline with Whisper, Gemma and ChromaDB

A four-person team built LectureLens (Mozhi), an offline study tool that turns lecture recordings and documents into summaries, key terms, quizzes and grounded Q&A with timestamp or page citations. The system runs entirely on a laptop using faster-whisper for timestamped transcription, Gemma via Ollama for generation, multilingual-e5-small sentence-transformers embeddings, ChromaDB as the vector store, and a Gradio interface, with no API keys or cloud services. Answers are restricted to the uploaded material and the tool states when a lecture does not cover a question rather than guessing.

LectureLens Mozhi turns a lecture recording or document into a study resource. You upload audio, a PDF, a Word file, a text file or an image, choose the language, and get a summary, key terms and a quiz. Then you can ask questions, and every answer comes only from your material and shows where it came from: a timestamp for audio, a page or section for documents. If the lecture doesn't cover the question, it says so instead of guessing. Everything runs locally on a laptop with open-source tools, with no API keys and no cloud. https://github.com/bavishnu11/hacktoberfest-hack-day-coimbatore-x-init-club-and-idea-club https://github.com/bavishnu11/hacktoberfest-hack-day-coimbatore-x-init-club-and-idea-club faster-whisper transcribes audio with timestamps. PDFs, Word files, text files and images are read into page- or section-labelled pieces. | Part | Tool | |---|---| | Speech to text | faster-whisper | | LLM | Gemma via Ollama | | Embeddings | sentence-transformers multilingual-e5-small | | Vector store | ChromaDB | | UI | Gradio | This runs fully open-source AI models locally—using faster-whisper for timestamped speech-to-text transcription and Gemma via Ollama for grounded Q&A, summaries, and quizzes. Managing local performance constraints when running compute-heavy open-source AI models faster-whisper and Gemma via Ollama and ensuring strict RAG grounding. Practical experience in building modular end-to-end RAG pipelines with vector databases ChromaDB , optimizing local multilingual embeddings, and structuring prompt constraints for accurate JSON quiz generation and local LLM orchestration. | Member | Contribution | |---|---| | Sanjay Vijay | faster-whisper transcription, timestamps, language setting, transcript cleanup | | Sujithbabu S S | Chunking, embeddings, ChromaDB, search function with timestamp metadata | | Vishal P | Ollama setup, summary, key terms, quiz, Q&A prompt, multilingual answers | | S J Bavishnu | Gradio app, repo and Git workflow, README, demo video, deployment, pitch |