MULTIMODAL HUMAN-SAFETY INTELLIGENCE

Evidence that
earns attention.

A private, production-oriented research platform that turns video, pose, temporal context, and optional audio into reviewable safety events.

Read the architectureInspect the evidence
93.75%best internal validation snapshot
<5 mspose path on GPU-class hardware
4×evidence streams designed to cooperate
5auditable event lifecycle states
SentinelFlow architecture map
VISUAL AUDIT

See the signal.
See the miss.

Open full gallery
True positive fall detections
TRUE POSITIVEfalling → fallen · confidence trace
True negative daily activity
TRUE NEGATIVEdaily activity rejected
Temporal alarm with pose overlay
RUNTIME ALARMpose + world + future warning
False positive hard case
HARD NEGATIVEsit / lie transition
False negative hard case
FAILURE ANALYSISocclusion / ambiguity
THE THESIS

Models observe.
Policy protects.

SentinelFlow separates perception from action. A detector can produce an observation; only the policy layer can create an operator-visible event. That separation makes confidence, degradation, acknowledgement, and audit explicit.

The public release exposes the system vocabulary and integration boundary while keeping private implementation, weights, and data outside the repository.