Physical AI, checked against measurements
Software that touches the physical world has to prove what it did. Everything we build follows four rules.
Four rules
- Measure
A command is not proof. Every physical action is checked against what the sensors saw, and an action that left no trace is flagged instead of trusted.
- Bound
Safety limits live in the device itself, not only in the cloud. A plan that would cross a limit is refused before it is sent, and the device refuses it again on its own.
- Learn locally
Every space is different. Models learn each room from its own measurements instead of assuming an average room.
- Beat the baseline
A model earns its place only by beating the plain rule it replaces, such as a fixed timer, on held-out data. If it doesn't, the plain rule stays.
Simulate, then bench, then rooms
- 1Simulation: the full software chain and the models run end to end before any hardware exists.
- 2Bench: each prototype stage tests only the unknowns that stage is for, on off-the-shelf parts where possible.
- 3Rooms: real spaces, measured, before any claim about what the product does.
Designs, firmware, models and test data are kept as source, so every result can be rebuilt and checked.