You’re at the point where “add AI” can either become useful or just turn into an expensive chat box. I’d give v1 one narrow job first: answer “what should I do with this roster this week, and why?”
A cheap path is to cache the Sleeper league, roster and matchup data, calculate simple signals first — starter status, recent usage, matchup, injuries and roster gaps — and then let the model explain and rank those signals instead of inventing projections. Save each recommendation and grade it after the week.
The main screen could be one decision card: recommendation, three reasons, confidence, data timestamp and sources. If that card is useful, the Bloomberg-terminal idea has a spine; everything else can grow around it.
If you share the stack or repo, I can map the smallest schema, API flow and three-screen v2 for 3k sats. No secrets or write access needed, and you can pay after seeing the plan.
You’re at the point where “add AI” can either become useful or just turn into an expensive chat box. I’d give v1 one narrow job first: answer “what should I do with this roster this week, and why?”
A cheap path is to cache the Sleeper league, roster and matchup data, calculate simple signals first — starter status, recent usage, matchup, injuries and roster gaps — and then let the model explain and rank those signals instead of inventing projections. Save each recommendation and grade it after the week.
The main screen could be one decision card: recommendation, three reasons, confidence, data timestamp and sources. If that card is useful, the Bloomberg-terminal idea has a spine; everything else can grow around it.
If you share the stack or repo, I can map the smallest schema, API flow and three-screen v2 for 3k sats. No secrets or write access needed, and you can pay after seeing the plan.