Live demo: win.mgdhia.com
Tell WIN a few things you loved — a movie, a show, a game, an anime, a book — and it tells you what to try next, from any of those categories, with a one-line reason for every pick.
WIN is a cross-category recommender built end to end: data pipelines from four public APIs, an evaluated retrieval engine (embeddings + keyword search with rank fusion), an LLM re-ranking step that was measured before it was shipped, a live A/B test, and a Next.js front end.
![]() |
![]() |
| Picks with reasons. "Like Interstellar, it is a thoughtful sci-fi drama…" | Your taste, your queue, and a taste card to share |
- Pick what you loved from 11,721 titles across 5 categories, search with poster results, or start from popular rows.
- Get 11 picks across categories, each with a reason. Picks show at once; an AI-refined list replaces them a few seconds later.
- Vibe mode: describe a mood ("cozy and a bit sad, for a rainy night") instead of titles.
- Live feedback: 👍 / 👎 reshape the list below the card you voted on.
- Next queue, taste profile, taste card, and Blend: send a link, get picks you'd both enjoy.
- Anonymous accounts (no sign-up), data saved server-side; a read-only share code is used for links.
flowchart TB
subgraph Ingest["Offline: import + embed"]
direction LR
A[TMDB · IGDB<br/>AniList · Open Library] --> B[Normalise<br/>shared genres, clean text]
B --> C[(PostgreSQL<br/>+ pgvector)]
B --> E[MiniLM embeddings]
E --> C
end
subgraph Request["Per request"]
direction TB
U[Liked items or<br/>mood text] --> V[Vector search<br/>pgvector, per liked item]
U --> K[Keyword search<br/>TF-IDF in memory]
V --> F[Reciprocal Rank Fusion<br/>+ popularity prior<br/>+ 👎 penalty]
K --> F
F --> D[Remove editions, sequels,<br/>same title, repeat creators]
D --> BAL[Balance across liked<br/>items and categories]
BAL --> R{LLM re-rank<br/>Claude Haiku 5.5}
R -->|ok| OUT[11 picks + reasons]
R -->|error / refusal| RULE[Rule-based reasons] --> OUT
end
C --> V
C --> K
- Retrieve. For each liked item, find its nearest neighbours (pgvector) and its best keyword matches (TF-IDF over genres, tags and descriptions). Each candidate is scored against its closest liked item, not an average — so "Death Note + Dune" gets picks for both, not a blur of the two.
- Fuse. Combine both rankings with Reciprocal Rank Fusion, plus a popularity prior (weight 0.3) and a penalty for anything close to a 👎.
- Clean up. Drop other editions and direct sequels of what you listed (cosine > 0.68, measured), the same title in another medium (the Dune movie when you listed the book), and more than one pick per author/studio. Balance the list across your liked items and categories.
- Re-rank. Claude Haiku 5.5 orders the 30 best candidates and writes the reasons. It can only answer with candidate indexes (structured output), so it can't invent titles. On any failure WIN keeps its own picks and rule-based reasons.
Every engine decision was tested offline on MovieLens first: real users, a time-based split (predict what each user liked later), and the same evaluation code for every model (evaluate.py, notebooks in notebooks/).
1 · Which model? (589 test users, full rating history)
| Model | NDCG@10 | |
|---|---|---|
| ALS collaborative filtering | 0.102 | best, but needs users who rated items in every category — a new cross-category app has none |
| MiniLM embeddings + popularity | 0.082 | ✔ chosen: works from descriptions alone, across categories |
| Popularity baseline | 0.079 | |
| Embeddings alone | 0.006 | recommend obscure titles nobody watches |
2 · Hybrid search (581 users, last 5 likes each, as in the app)
| Retrieval | Recall@30 | NDCG@10 |
|---|---|---|
| Vector only | 0.090 | 0.062 |
| Keyword only (TF-IDF) | 0.094 | 0.062 |
| Hybrid (RRF) | 0.111 (+24%) | 0.070 (+14%) |
Both gains are significant (paired bootstrap, 95% CI). A first run made hybrid look worse because it fused the weakest keyword setting — comparing methods at equal settings fixed that.
3 · Does the LLM help? (100 users, ~200 Claude calls, under $1)
| Method | Recall@10 | NDCG@10 | |
|---|---|---|---|
| Engine | 0.032 | 0.055 | |
| Engine + LLM re-ranking | 0.046 (+43%) | 0.046 | recall gain significant; NDCG change is noise |
| LLM alone | 0.046 | 0.062 | not significant, and 11.5% of its titles aren't in the catalog |
So WIN uses the LLM as a re-ranker over retrieved candidates, not as the recommender. Live, the A/B test (30% of new users get the engine only) compares 👍 / queue rates per card on the /stats page.
Other measured choices: popularity weight 0.3 (84% of the best accuracy with 1.7× the catalog coverage of 0.5), closest-liked-item scoring instead of averaging, keeping the title in the embedded text (removing it cost 19% NDCG).
| Layer | Tools |
|---|---|
| Data | TMDB, IGDB (Twitch), AniList, Open Library importers · PostgreSQL 17 + pgvector |
| ML / search | sentence-transformers (all-MiniLM-L6-v2) · scikit-learn TF-IDF · Reciprocal Rank Fusion · implicit (ALS, for evaluation) |
| LLM | Claude Haiku 5.5 via the Anthropic SDK, structured output, fallback + pause on errors; cost-protected by a daily call cap, a per-user/IP rate limit and server-side A/B arm checks |
| API | FastAPI · Pydantic · psycopg 3 |
| Front end | Next.js 15 · React 19 · Tailwind CSS 4 · server-rendered taste-card images (next/og) |
| Evaluation | pandas · NumPy · Jupyter · paired bootstrap · Wilson intervals for the live A/B test |
| Quality | ruff · TypeScript |
| Ops | Docker Compose · uv |
cp .env.example .env # add free TMDB + IGDB keys; ANTHROPIC_API_KEY is optional
docker compose up -d --build # database :5433, API :8000, web app :3000
cd backend && uv sync
uv run python -m app.ingest anilist --limit 2000 # anime, no key needed
uv run python -m app.ingest openlibrary --limit 3000 # books, no key needed (~10 min)
uv run python -m app.ingest tmdb --limit 5000 # movies + TV
uv run python -m app.ingest igdb --limit 3000 # games
uv run python -m app.embed # vectors for new or changed itemsOpen http://localhost:3000. Without ANTHROPIC_API_KEY, WIN runs on its own engine and rule-based reasons.
Experiments
cd backend
uv run python -m app.recsys.compare [--tmdb] # model comparison
uv run python -m app.recsys.hybrid_eval # hybrid search experiment
uv run python -m app.recsys.llm_eval --users 100 # LLM experiment (needs ANTHROPIC_API_KEY)backend/app/
ingest/ importers (one module per source) + shared genre mapping
embed.py embedding pipeline keyword.py TF-IDF index
recommend.py the engine (retrieve, fuse, filter, balance, explain)
rerank.py LLM re-ranking users.py / share.py / metrics.py accounts, Blend, A/B metrics
recsys/ offline evaluation, model comparison, hybrid + LLM experiments
frontend/ Next.js app and components
notebooks/ EDA, model comparison, every experiment above with its results
Movies and TV from TMDB, games from IGDB, anime from AniList, books from Open Library; offline evaluation on MovieLens (GroupLens Research). This product uses the TMDB API but is not endorsed or certified by TMDB.


