One call — POST /v1/score, before you invoke the model — returns a calibrated 0–1 risk score, a tier, and up to three human-readable reasons. It scores across three layers built from the account's own history:
The three layers combine into one calibrated verdict, not a pile of raw signals — an account that trips cost, network, and behavior at once scores far higher than one with a single odd spike. That keeps false positives off your real power users while the genuine reseller rises to the top of the queue.
- Cost & consumption
- Token-velocity bursts, robotic timing, instant-first-call after signup, cheap-model concentration, tiny-prompt→huge-completion ratios, 24/7 activity.
- Network
- Datacenter-IP origin (AWS/GCP/Azure/OVH/Hetzner/…), IP rotation, and sybil clusters (many accounts, one IP).
- Behavior
- Inter-request gap distributions, session fragmentation, day-of-week diversity (always-on automation vs. human diurnal rhythm).
Capabilities
Calibrated, explainable scores. Every verdict ships with its top reason codes
Optional per-customer LightGBM model, trained nightly on your labels
~30 ms median latency; industry-tuned profiles (gateway, coding, chatbot, consumer)