RecallAI

Re-reading notes is passive and slow. RecallAI turns uploaded study material into practice: a retrieval-augmented generation pipeline indexes the documents and generates quizzes and flashcards grounded in what the student actually uploaded.

• Next.js and React application deployed on Vercel.
• Firebase for storage and data.
• OpenAI and LangChain for the generation pipeline; Sentence Transformers for document embeddings.
• Clerk for authentication and Stripe for payments.

• Retrieval-augmented generation over fine-tuning. Questions stay grounded in the uploaded documents and new material works immediately; the trade-off is that answer quality depends on retrieval quality.
• Managed services (Firebase, Clerk, Stripe, Vercel) instead of self-hosting. Fast to ship and nothing to operate, in exchange for vendor lock-in.

• Identity and payments are delegated: Clerk holds sessions and credentials and Stripe holds card data, so the application never stores either.
• Uploaded documents and the generated quizzes and flashcards live in Firebase, scoped to the signed-in user.
• Cost scales with use rather than with time: Vercel and Firebase have no idle infrastructure to pay for, and OpenAI is billed per generation.
• The gap is measurement. Retrieval and generation changes were judged by hand; an evaluation set of documents with expected questions is the first thing I'd add.

• Build the evaluation set before the prompt. A fixed corpus with expected questions and answers turns every retrieval or generation change into a score instead of an opinion.
• Version the chunking and embedding choices. Chunk size, overlap, and the embedding model are product decisions in a RAG app; recording them with the index makes a change comparable and reversible.
• Keep the thin-interface discipline. Delegating auth, payments, and storage to Clerk, Stripe, and Firebase was right for speed; wrapping each behind a small module would make the lock-in a choice rather than a fact.