Senior Researcher, AI Transformation @ Rainbow Robotics I lead AX for the RBQ quadruped: turning how we develop, test, and ship robots into workflows where AI agents do the heavy lifting.
- One app, 7 platforms, 3 stores — rebuilt the RBQ controller with AI agents: one TypeScript codebase for Android · iOS · Windows · Linux (x86/ARM) · Web · Steam Deck, live on Google Play · App Store · Microsoft Store
- Open-sourced it — published daily as Apache-2.0 source, behind an automated gate that stops the publish if anything internal slips in
- AI-native pipeline — AI review on every PR, labeled issues turned into AI-written fix PRs, nightly multi-platform builds pushed to the stores (ARM build 6 h → 23 min, PR checks 2 h → 8 min)
- Simulation as the safety net — 38 MuJoCo scenarios run headless every night after the build, SDK wire compatibility checked against the last release; RBQ-10 submitted to MuJoCo Menagerie
- Robot knowledge AI — RAG over code, docs, manuals, and change history: answers in ~6 s, 97–99% on a 102-question eval, used by the team and by the review and issue agents
- Agents that touch the robot — MCP servers (24 robot tools, 10 app tools) let AI drive the robot, simulator, and app, and walk every app screen without a robot
- Reliability from field logs — trade-show logs turned into 28 fixes: stop after stick release 3.92 s → 0.72 s, camera frame loss 15–22% → 0
- Language → behavior (prototype) — an LLM agent docks the robot in simulation through function calls; an in-app AI assistant is in test builds
Before robots: 8 years in games, XR, and digital twins — 4 games and 3 apps shipped, 12-month projects done in 1–3 months with AI.
| RBQ controller | Open-source quadruped controller app — React Native + Electron, 7 platforms, 3 stores (at Rainbow Robotics) |
| task-board | One board for many parallel Claude Code sessions — questions and reports per task, [Send] wakes the session |
| context-forge | Bootstrap for AI-augmented repos — 82 curated sources, auto-synced |
| graph-rag-viewer | Hybrid RAG + MCP server — 9,556-node knowledge graph, 100% hit rate |
| harness-bench | CLI that scores how AI-native your Claude Code setup is (8 axes) |
| claude-mux | PTY multiplexer that drives interactive Claude Code sessions instead of claude -p |
| Molten | Terminal wrapper built for AI coding agents (Tauri · Rust) |
| AI_Language | Compressed agent-to-agent protocol — 40.7% fewer tokens |
| Data2Avatar | CCTV → digital twin safety monitoring, 4 pose models, ≤150 ms sync |
| VTS Visualizer | Industrial 3D digital twin — 12 months → 1 month with AI |
More on the portfolio.
LLM Agents RAG MCP Claude Code · ROS 2 DDS WebRTC C++ MuJoCo · React Native Electron TypeScript Python · GitHub Actions Docker · Unity/C#