03 / Across the stack

A robotics engineer who works every layer of the robot.

Robots fail at the seams: a calibration that drifts, a ROS node that drops frames, a model that ships but can't be retrained, a platform nobody can query. I've worked on every layer, on real robots in production, and I design each one with the next in mind.

  1. L4

    Platform & infrastructure

    The data platforms, pipelines and training infrastructure that turn robot logs into models, and models back into robots.

    • Led data infrastructure and platform work at Boston Dynamics: PB-scale data for 50+ robot-learning researchers
    • On-robot model deployment with ONNX and TensorRT, serving 1M+ inferences a day
    • Forge · Tessera · CoRL 2026
  2. L3

    Perception & robot learning

    Models that see and act: detection and segmentation running on robots, and the VLA and imitation-learning policies on top.

    • Semi-automated annotation with SAM, CLIP and active learning for manipulation supervision
    • YOLO and U-Net perception for people, robots and forklifts on Fetch and 6RS fleets
    • Onboard VLM/VLA hill-climbing at Cruise: data mixes, ablations, agentic labeling
  3. L2

    Robot software

    ROS systems that hold up in production: correct QoS, lifecycle nodes, bounded callbacks, recording and bringup.

    • Perception nodes running on 1,000+ deployed warehouse robots
    • ROS Time Machine: event-triggered MCAP recording in C++17 / ROS 2
    • Robotics Agent Skills: production ROS patterns for AI coding agents
  4. L1

    Sensors & calibration

    Clean, trustworthy signals: cameras, depth, LiDAR and IMUs, calibrated and fused.

    • Point-cloud camera calibration that cut site setup from 4 hours to 15 minutes (patent filed)
    • Monocular SLAM scale from ground planes plus CAN bus and IMU fusion at Panasonic
    • Anvil: catch scene and camera drift between collection sessions