"See what you're spending on AI, and enforce a cap in-process."
Sentinel reports GPU utilization and cost-accrual metadata in real time. Monitoring-only and metadata-only — cost enrichment against provider rates happens downstream on the platform.
The LUT Agent SDK enforces a spend cap inside your application process — raise, warn, callback, or SIGKILL the process on breach. This is process-level enforcement, not cluster-level.
Track accumulated spend and remaining budget programmatically through the LUT Agent SDK's client state — per session, per process.
The Tier 0 measurement CLI (pip install lutflow) reads a vLLM server's own Prometheus metrics to compare self-hosted compute cost against API pricing you provide — no cluster, no account, no GPU sensor required.
Set a budget policy per process via the SDK. Enforcement is deterministic within that process: the configured strategy fires exactly at the configured cap.
PCPO is a heuristic cost estimator calibrated with LMSYS percentiles (p50) — it does not ship a trained model artifact. Use it as a starting reference, not a guarantee.