Persistent World Models
Mihawk turns the physical infrastructure you already have into a persistent world model — one that tracks who was where, when, across imperfect and changing environments. No new hardware. No data leaves the premises.
The Problem
Traditional
Stateless
Per-frame accuracy doesn't produce understanding. Each moment is treated as independent — no memory, no history, no context.
Mihawk
Persistent
A structured, time-indexed record of what actually happened — grounded in space and time, queryable, and constrained by recorded evidence.
Alerts need speed. Identity needs time. Real understanding needs both.
The Approach
Structure from imperfect observations
Observations from existing sensors are fused into a shared, persistent model of the environment. Positions, relationships, and events are grounded in space and time — so the system knows where things are, even across viewpoints, occlusions, and changing layouts.
Memory over time
People and objects are tracked across views and over time into a searchable record of who was where, when. Entity continuity, trajectory history, event recall — all persistent and queryable.
Inference under constraints
Reasoning agents build and test hypotheses over the model, ruling out what physical and spatial constraints make impossible. Hypothesis assembly, scenario replay, counterfactual check.
Live Deployments
Each deployment is a different query over the same underlying world model. Persistent situational understanding is the capability.
Education
From a single existing classroom camera, Mihawk maintains a persistent record of who was present and when — despite occlusion, natural movement, and imperfect camera placement.
Security
Entry and exit decisions resolved with identity, policy, and event context on top of existing access infrastructure. Real-time authorization with full audit trail.
Traffic
Vehicle movement, violations, and evidence reconstructed from existing roadside cameras — the same world model applied to vehicles, pedestrians, and infrastructure.
Research
The principles behind the product are formalized and benchmarked in our research — not asserted.
Published
Grey-Box Port-Hamiltonian State-Space Models with Recoverable Physical Structure
~10× smaller and ~10⁴× more accurate than Transformers across 13 SciML benchmarks.
Paper →Preprint
Compositional Port-Hamiltonian World Models for Structured Dynamics Transfer
~10× more stable under robot-embodiment swap. Robust under topology and sensing change.
Preprint →Design Philosophy
"To reason about the physical world at scale, a system must not only capture structure — it must think with it."
Mihawk understands a space through how things relate and persist — not through perfect measurement — so it holds up on real, imperfect deployments.
Cameras move, sensors are added, layouts change. Mihawk folds new and imperfect signals into the same persistent world model without rebuilding from scratch.
Not a perfect simulator — just enough structure to constrain what's possible and compare candidate futures, degrading gracefully rather than collapsing.
If you operate in environments where context, accountability, and time matter — or if you are building in physical-world intelligence — we'd like to hear from you.