Skip to content

DEPLOY.md

Latest commit

 

History

History
49 lines (44 loc) · 3.39 KB

DEPLOY.md

File metadata and controls

49 lines (44 loc) · 3.39 KB
Raw

Deploying GoalForge on Vercel

The current (statistical) model is tiny (api/model.json ≈ 230 KB — 48 teams, ~1250 players) and inference is analytic, so the whole app runs on Vercel: a static frontend + tiny Python serverless functions. This mirrors the working reference project aevum-orrin/applet-material.

Structure

  • public/ — static frontend (outputDirectory in vercel.json), served at /.
  • api/health.py, api/teams.py, api/squad.py, api/predict.py — one serverless function per endpoint, hit directly at /api/<name>. Python standard library only (no numpy/fastapi → nothing to install, cannot fail to import). They share api/_engine.py (analytic Poisson inference: scoreline + per-player Poisson-thinning anytime probabilities) and api/model.json.
  • vercel.json — outputDirectory: public; a negative-lookahead rewrite sends every non-/api/ path to the frontend; includeFiles: "api/**" bundles the whole api/ dir (the shared _engine.py and model.json) into each function's lambda. Each api/*.py is a separate lambda, so every one needs _engine.py shipped with it; each handler also does sys.path.insert(0, dirname(__file__)) so import _engine resolves inside the lambda.
  • requirements.txt — empty (the functions need no third-party deps).
  • .vercelignore — excludes pyproject.toml (and src/). This matters: newer Vercel builders scan pyproject.toml, and ours lists fastapi (optional-deps) while the real FastAPI app lives in src/goalforge/api/app.py. If Vercel sees that, it switches to "FastAPI backend" mode, hunts for a single ASGI entrypoint, finds only the handler functions, and fails the build with "No FastAPI entrypoint found". Ignoring pyproject.toml/src/ keeps it on the classic static-site + per-file Python serverless-functions path (what applet-material uses).

Connect (you handle the GitHub↔Vercel link)

  1. Push to GitHub (done).
  2. vercel.com → New Project → import GoalForge-AI. Framework preset Other; root = repo root.
  3. Deploy → public URL. Every git push redeploys. Check <url>/api/health first — it returns {"status":"ok","teams":48} when the model is bundled.

Updating the model / squads (real 2026 World Cup)

python scripts/train.py               # (re)fit the Dixon-Coles team layer -> models/agent_intl.pkl
python scripts/scrape_wc2026.py       # scrape the 48 real 2026 squads (Wikipedia) -> scratch cache
python scripts/fetch_understat.py     # club xG/xA per player (5 big leagues; free Understat mirror)
python scripts/build_player_form.py   # name-match club form onto the 2026 squads -> player_form
python scripts/build_wc2026_model.py  # combine -> api/model.json (xA assists + club-enriched scoring)
git add api/model.json && git commit -m "update 2026 model" && git push

build_wc2026_model.py reuses the validated team checkpoint for the scoreline, and for the ~1/3 of squad players with top-5-league data blends recent club xG (scorer) and xA (assister, the validated chance-creation signal); the rest fall back to international goals/caps + a position prior. See docs/evaluation.md §8-11 for the honest evaluation behind these choices.

Endgame (after Phase 3)

Once neural models arrive, the heavy model moves to a Hugging Face Space (torch + GPU); Vercel keeps serving the frontend, calling the HF backend. See docs/workflow.md.