Physical AI Midlayer

active

Intent-based robotics R&D platform: agentic multi-agent orchestration running offline on NVIDIA Jetson, bridged to ROS2 through MCP.

Oct 2025 – Present · @ Device Insight GmbH · closed source

Python
LangGraph
Microsoft Agent Framework
MCP
vLLM
NVIDIA Jetson
TensorRT
quantization
ROS2

What it is

An R&D platform for intent-based robotics, built for the CTO Office of a leading robotics manufacturer: robots that act on high-level task intent, orchestrated by agents running offline on the robot.

What I did

Led the R&D for the platform's agentic-orchestration layer: scoped open briefs, built the prototypes and drove the architecture decisions.
Built a multi-agent orchestration prototype on NVIDIA Jetson after evaluating LangGraph and Microsoft Agent Framework hands-on, with on-device inference (vLLM, quantization) and error-tolerant task execution.
Built ROS2 to MCP interop and validated it on real hardware with a five-finger robotic hand.

Outcome

An 80-page whitepaper I authored set which components advance to next-stage research and which architectural direction the platform takes.