Attributes LLM token spend to teams, features and prompt versions from request traces, pricing each request at the rate in force when it was made. Finds a prompt version that tripled one team's tokens at robust z 10.17 where a classical z-score reads 2.90 and measures the 7.2% false-hit rate a prompt cache would carry.
python sql lsh data-engineering minhash observability anomaly-detection finops robust-statistics duckdb showback llmops budget-forecasting cost-attribution
-
Updated
Aug 19, 2026 - Python