How It Works
Our engines combine market data with contextual signals via AI tool-calling. We currently support NFL (winners, ATS, totals) and Stocks (plays with entries/targets/confidence). Results are settled and tracked automatically.
Architecture
- Provider adapters: The Odds API (lines), ESPN (injuries + recent results).
- AI orchestrator: Tool-calling loop with deterministic prompt policy.
- Persistence: One prediction per game/version; odds snapshot stored.
- Settling: Post‑kickoff one‑off checks to mark actuals and hits.
Data Sources
- Odds (The Odds API): spreads, totals, moneylines; consensus snapshot is saved on ingest.
- Injuries (ESPN): team injury pages; we summarize key positions (QB/OL/DL/CB/S).
- Recent form (ESPN): last four weeks of games, merged with scores.
Tools & Prompt Policy
For each unplayed game:
get_vegas_odds: spreads/totals/moneylines.get_recent_games: past 4 weeks per team (W/L, opponents, scores).get_injuries: current injuries, highlighting key positions.
The model is instructed to weigh multiple signals and not pick favorites purely from the spread. Output includes home/away scores, winner, win probability, confidence, and rationale.
ATS & Totals Logic
- ATS: home covers if
home_score + home_spread > away_score (push on equality).
- Totals: compare predicted/actual totals to market total; label OVER/UNDER/PUSH.
Settling & Tracking
- We schedule a few post‑kickoff checks (e.g., +4h, +8h, +24h). No constant polling.
- Once final, we store actual scores and mark ML/ATS/Total as hit/miss/push.
- Weekly groups and badges make scanning performance easy; more charts coming soon.
Accuracy & Limitations
- Market lines can move; we snapshot consensus but do not stream live updates.
- Injuries and recency are scraped; format changes may require parser tweaks.
- AI is guided by tools and policy, but can still deviate — rationale is included for transparency.
Access
- Signed‑in users can view predictions.
- Paid tiers (coming soon) will unlock real‑time updates and deeper breakdowns.