Popcorn Rec
A movie and TV discovery app that turns "what should we watch?" into a two-second answer. Built on Next.js 16 and TMDB, with a two-layer recommender that ends in an AI curator constrained by a Zod-validated tool schema — so it can only return real titles, and degrades gracefully when the model is unavailable.
Built on Next.js 16 and the TMDB catalog, it slices thousands of titles the way people actually think - not just by genre, but by mood ("Date Night," "Mind-Bending," "Spine-Chilling") and by era (the neon 80s through the 2020s).
The heart of it is a two-layer recommender. The first layer analyzes what you've liked and pulls related titles - but that's a firehose of 50 technically-relevant results with no point of view. So I added an AI curator on top: it takes your taste profile and picks the ten best, each with a one-sentence why that references the movies you already loved. It reads like a friend's text, not a spreadsheet.
Key Decisions
The AI is a reliable component, not a demo. It answers through a typed, Zod-validated tool schema, so it can only return real titles - no hallucinations, no broken JSON. It runs on Claude in the cloud or a local model via Ollama, with a graceful fallback if the AI layer is down.