Proof you can run, not slides you can read — live demos, ventures in active development, and the frameworks behind the pilot-to-production method.
Three demos running real inference, open to anyone. Each page states exactly what the demo does and does not prove.
Upload a photo of a plant, warehouse or lab — the model surfaces automation candidates and safety observations.
Open the demoDrop a photo of a single part — the model narrates what an industrial inspection layer would flag.
Open the demoAsk plain-language questions about the EU AI Act and the AI Pact — answers grounded in a curated corpus with inline citations.
Open the demoQwen3-30B-A3B at 19.3 t/s · GPT-OSS-20B at 22.8 t/s — measured on the lab's Strix Halo bench, single runs, no tuning.
Field notes from the benchThe in-house venture portfolio, in active development — the runnable record behind the method. Every card states its maturity; none is in production.
The coordination plane for physical AI
Exercises: IntegrationReliability
AI-Powered Product Management Intelligence
Exercises: ArchitectureIntegration
AI-Powered VC Intelligence
Exercises: DataArchitecture
Vehicle Intelligence Platform
Exercises: IntegrationArchitecture
Creative AI Super-Assistant
Exercises: ArchitectureIntegration
The Regulatory Shield
Exercises: GovernanceData
AI-Powered Business OS
Exercises: ArchitectureData
Static Analysis for AI-Generated Code
Exercises: GovernanceReliability
AI Adoption Copilot for European SMEs
Exercises: GovernanceData
AI-Powered Enterprise Transformation Intelligence
Exercises: GovernanceArchitecture
How the complete system works
Select a stage to follow the system end to end.
01 / 08Device & Control
Sensors & environment. The physical world, sensed — cameras, depth, radar and OT signals.
See how we engineer this02 / 08Embedded Platform
Embedded platform. A real-time, deterministic, securely-updatable edge runtime.
See how we engineer this03 / 08Perception & Robotics
Perception. Mapping, detection and world-state resolved on-device.
See how we engineer this04 / 08Intelligence & Knowledge
Intelligence. Open-weight models and VLM/VLA policies propose an action.
See how we engineer this05 / 08Device & Control
Deterministic control. A deterministic controller turns a proposal into a precise command.
See how we engineer this06 / 08Trust, Safety & Cybersecurity
Safety & cybersecurity. An independent boundary accepts or blocks every command.
See how we engineer this07 / 08Perception & Robotics
Physical action. The validated command drives the machine in the real world.
See how we engineer this08 / 08Orchestration & Operations
Operations & fleet. Telemetry returns to operations; the system improves across a fleet.
See how we engineer thisThe method, written down — frameworks and self-serve tools an engagement runs on.
The reference architecture behind every engagement — published as an arXiv preprint.
Map deployment constraints to a hardware platform class in a few questions.
Classify an AI system into the Act's four risk categories.
A first-pass payback model for a robotics cell.
A self-check across the five dimensions a Production Readiness Review assesses.
Thirty minutes with the person who runs this lab.