What's real, what's mocked

A plain-English tour of everything behind SportAI, and exactly which parts of the product you're seeing on this public demo site.

The 30-second version

SportAI is a platform for college-basketball-and-beyond analytics. It has three working pillars today: game predictions (which team will cover the spread, which total will hit), draft scouting (which NBA prospects are undervalued), and biomechanics (which players are showing injury-risk movement patterns on film). Everything is tied together by an internal team wiki, a receipts page that keeps the models honest, and an automated research loop that retrains models nightly.

What you can click right now

Section	Status	What it shows
Dashboard	Demo data	The real NCAA prediction UI — fully interactive, backed by a small mock API Worker that serves sample games and prospects.
Draft (False Floor)	Demo data	Real big-board and compare screens, hitting the same mock Worker with a sample 2025-class prospect list.
Team Wiki	Full content	All 292 D1 team profiles, auto-compiled from the production wiki markdown.
Receipts	Live	Cloudflare Pages site that already tracks every model prediction vs. the actual outcome.
Biomechanics	Preview page	Static explainer — the full pipeline needs GPU + video and is runnable on the workstation.

What lives behind the curtain (not on this demo)

Live FastAPI backend — runs on localhost:8888 on the founder's Mac. Serves real predictions and draft inference.
XGBoost models — spread, totals classifier, totals regressor, and a draft-likelihood model, all trained locally and tracked in an experiment log.
Supabase auth + Postgres — local first (port 5433), with email/magic-link auth already wired.
n8n workflow automation — nightly scraping, model retraining triggers, content publishing hooks.
Biomech pipeline — YOLOv8-pose, TimescaleDB, Prometheus alerts. GPU-bound.
React Native mobile app — Expo scaffold, not yet published to TestFlight.
Autoresearch loop — scheduled job that scores new features, runs training, and writes a markdown report each morning.

How the pieces fit together

Think of it as three concentric circles. The inner circle is the data and the models — game histories, player stats, pose skeletons, trained XGBoost binaries. These stay on proprietary storage (UNAS + local SSD) and never leave the founder's control. The middle circle is the API layer — FastAPI routers that expose predictions, watchlists, and draft scouts, plus the False Floor read-only adapter that lets external apps consume draft data without touching the main DB. The outer circle is what you see in this demo — the dashboard, the wiki, and the receipts page, which are designed to be shareable and auth-gated separately.

Where it's going

Near-term: Cloudflare Tunnel for authenticated access to the live backend from anywhere, so first pilot clients can use the real models (not the mock).
Next: Supabase-based multi-tenancy (per-client row-level security already scaffolded), Stripe for three-tier subscriptions ($49 / $199 / Custom).
Later: Move inference API to a VPS or Cloudflare Workers + R2 + D1, while keeping training and proprietary data on the home lab rig.
Parallel track: Finish TestFlight build of the mobile companion app; publish the biomech trainer-facing dashboard as a separate product tier.

If you want to go deeper: the comprehensive project breakdown lives at backend/autoresearch/Comprehensive_Project_Breakdown.md inside the repo. For anything demo-related, use the hub home — every section has a working back link.