Mihawk

Organizations operating physical spaces such as classrooms, secured buildings, and roads rely on cameras and sensors that can see but don't remember. Traditional vision systems are stateless and per-frame, producing alerts and detections but not a persistent understanding of who or what was where over time. This forces a tradeoff between systems built for speed (alerts) and systems built for identity and accountability, with neither providing real situational understanding. Per-frame accuracy alone does not produce understanding of how spaces, people, vehicles, and events evolve.

Mihawk turns existing physical infrastructure, like cameras already installed, into a persistent world model that is queryable, reasoned over, simulated, and acted on. It works in three stages: perceiving structure from imperfect, heterogeneous observations and fusing them into a shared spatiotemporal model; remembering by tracking entities across views and time into a searchable record of who was where and when; and reasoning by building and testing hypotheses against the model, ruling out what spatial and physical constraints make impossible. The system is designed to be relational rather than dependent on perfect measurement, and to absorb new or imperfect signals from changing infrastructure without requiring redeployment. The same underlying world model is applied across different use cases, including classroom attendance tracking, access control decisions, and reconstructing vehicle violations from roadside cameras.