Operating an AI-native product with cost intelligence built in.

The product

The product creates AI synopses and the most important things to watch across teams, initiates an AI-analyzed brief for each team, presents team statistics, and provides conversational answers to user questions.

Its fantasy feature set supports drafts and ongoing start-or-sit and trade analysis.

The operating challenge

Useful analysis across many teams requires timely information, observable model behavior, and manageable operating cost. AI economics therefore became part of the product architecture.

The telemetry system

CalSportsAI records each completed AI generation as structured product telemetry. The system captures pipeline and operation, model and provider, request and run identifiers, user and team context, token consumption, web searches, estimated cost, latency, and generation metadata.

• Daily and per-run operating views
• Input, output, and cached token accounting
• Cost attribution by feature and operation
• Context linking requests to users, teams, and leagues

What the evidence means

Across the approved 12-day sample, the product recorded 454 API calls, 6,623,097 input tokens, 465,545 output tokens, 316 web searches, and $29.4414 in estimated model cost—an average of $2.4534 per day and $0.0648 per call.

The recorded dataset contained no failed calls. Because telemetry is primarily written after completed provider responses, that result should not be interpreted as proof of perfect system reliability.