Tastemake
Tastemake is an experimental recommendation product that helps people understand and refine their taste across books, movies, TV, and games.
Instead of only asking whether someone liked an item, Tastemake also builds a visible Taste Profile that explains the patterns it thinks are shaping those reactions. Users can correct those patterns, narrow them, or reject them, and those corrections influence what happens next.
Current product loop
Favorites -> Recommendations -> React -> Taste Profile learns -> Better recommendations
The first-run flow stays intentionally light:
- add a few Favorites
- get recommendations
- react with Loved, Liked, Didn't like, Saved, or Not interested
- refine deeper taste details only when useful
- inspect what Tastemake learned in the Taste Profile
The product keeps experienced taste separate from future intent. Saving or browsing does not automatically become evidence that a user likes something.
AI and product boundaries
Tastemake uses live AI in two places:
- Recommendation ranking and explanation
- Taste Profile hypothesis generation
The model does not control the whole system. Software still decides:
- what counts as taste evidence
- which catalog candidates are eligible
- whether model output passes validation
- what user corrections take priority
If a live recommendation call fails or is rejected, Tastemake can fall back to catalog-based recommendations rather than presenting weak model output as trustworthy.
Catalog and data
Tastemake works with real catalog data across:
Provider data is normalized into Tastemake's own catalog shape, with a canonical store used to improve reuse and metadata consistency over time.
Quality and testing
The repository includes automated coverage for:
- recommendation behavior and candidate quality
- AI output validation
- evidence and intent rules
- Taste Profile corrections
- persistence and export
- desktop and mobile layout
- accessibility
- end-to-end product flows
- recommendation evals
Run the main test suite with:
npm test
Run locally
From the repository root:
python3 -m http.server 4173
Then open http://localhost:4173.
Some live AI and catalog features depend on server-side environment configuration, so the deployed Vercel app is the best way to try the full experience.
Research and product notes
The experiments/ folder contains taste-model and recommendation experiments that informed the product. The notes/ and docs/ folders capture product decisions, QA findings, and implementation context.
Deployment
Vercel builds a preview deployment for each pull request. Merging to main deploys production (https://tastemake.vercel.app), so merges to main happen only after review. For contributors and new sessions, docs/HANDOFF.md is the place to start.
