The questions we haven’t settled yet.
A model of an operation can be right on Monday and wrong by Friday — sources disagree, entities merge, the world moves underneath it. That gap is where our research lives. It isn’t a lab off to the side: every direction feeds back into the systems we ship, or it doesn’t continue.
Six directions. Each one feeds the same systems.
Knowledge graphs
Keeping a model of an operation correct as reality shifts underneath it — resolving entities, reconciling conflicting sources, and expiring stale relationships without losing history.
Spatial computing
Reasoning about movement, geometry, and place over time, so position and trajectory become queryable facts rather than raw coordinates.
Agentic systems
Giving long-running agents reliable plans, tool use, and self-verification, so they can act and check their own work against ground truth.
Computer vision
Turning visual streams into structured entities and events that link back to a knowledge model, under real-world noise and occlusion.
Edge inference
Running reasoning close to where signals originate, trading model size, latency, and connectivity so decisions hold up when the network doesn’t.
Reasoning systems
Combining learned models with explicit structure so answers can be traced, constrained, and defended — not just generated.
Grounded in the problem, shipped into production.
We start where the friction is real — an operational problem someone is living with — and work backward to the method that solves it.
- Every project starts from a real operational problem, not a benchmark.
- Work ships into a production system, or it isn’t done.
- Results are held to ground truth, not to how good the demo looks.
Questions the field hasn’t settled.
The hard, unfinished parts — shared across the field, not solved by anyone. We state them plainly, because an honest question makes for better work than a confident answer.
Staying current
How does a knowledge model stay correct when its sources disagree and the world keeps changing beneath it?
Trustworthy autonomy
How far can an agent act on its own before a person needs to step in — and how does it know where that line is?
Intelligence at the edge
How much reasoning can move to a constrained device without losing the context that lives in the central model?