How I measure what I ship, and how the AI actually works.

The eight AI features

• Draft grades — Post-auction team grade (A–F) with rationale across category strength, risk, upside, depth
• Trade analyzer — Evaluates proposed trades for both sides, flags category gaps, surfaces fair-value read
• Keeper recommender — Given salary inflation, ages, position scarcity, which keepers hold value?
• Waiver advisor — Ranks available players against your roster holes, not against raw projection
• Weekly insights — Per-team narrative of the week's moves and standings movement
• Bid advisor — During live auction: suggested max bid based on remaining budget + positional needs
• League digest — Weekly AI-generated recap with team grades and trade polls
• Insights cache — Shared cross-league cache to avoid duplicate inference for identical prompts

How the cache works

League-context prompts are large and regenerating them every page view would be expensive and slow. Instead, the MCP MLB proxy caches MLB data in SQLite (shared across all leagues); AI responses are cached per (team_id, feature_key, data_hash) tuple so owners see near-instant responses on repeat views.

Tracking now

• Pageviews — Page-level tracking with SPA-aware navigation events
• User identity — Authenticated users identified by email for session continuity
• Feature adoption — Auth flows, trade proposals, waiver claims, keeper saves, watchlist actions
• Auction engagement — Nominate, bid, proxy bid, chat, init, finish, force assign, watchlist toggle, WebSocket reconnects
• Error tracking — React error boundaries report crashes with component name and stack trace

Planned

• Page performance — Load times, slow API calls, rendering bottlenecks
• Click tracking — Granular interaction events: tab switches, modal opens, sort changes