Robotics and AI engineer. I run a self-funded lab in London and Kuala Lumpur where USD 1k arms learn the jobs of USD 20k ones, and I build the agents that run companies. Before that, founding AI engineer at Goldman Sachs and Fern Labs (acq. by Poolside).
01 · Robots
One pipeline: record a real task, rebuild the place in simulation, train pi0.5, SmolVLA and ACT policies there, then take them back to the arm. Six rooms so far. Every clip below is from my own runs, with what worked and what didn't.
Home lab · Sep 2026 –
An XLeRobot: two SO-101 arms on a wheeled base with three cameras. I took it from loose servos to leader–follower teleop, built a dual-arm demo recorder, and wrote my own cloth-folding training stack.
No autonomous real fold yet. The first policy trained for 1,000 steps; the next run trains longer on the cleanest demos.
Cloth folding · Isaac Sim
I reproduced the LeHome Challenge folding policy on rented GPUs, then went after a harder fold: lift a side so gravity lays the sleeve underneath.
Home Twin · Oct 2026
Film a room, rebuild it as a 3D map and a Gaussian splat, drop it into Isaac Sim, and let a pretrained pi0.5 policy act in it. Then ask for a chore in chat: Claude picks the spot, starts the GPU and sends back the video.
Restroom cleaning · MuJoCo, Isaac, DimOS
Claude looks at the cameras and names the next step; a SmolVLA trained on the teacher's step labels does the scrubbing. I also ran PPO practice and DAgger, and wired the same skills into DimOS over MCP.
Farm · MuJoCo
Tomato vines grow, ripen and get lowered on a real crop schedule, and the robot harvests twice a week. I recorded harvest demos and trained SmolVLA and ACT to pick from cameras alone.
Forest · DimOS
A procedural forest in DimOS where a simulated Unitree Go2 walks to soil-sample sites. One MCP call starts the survey; an agent runs the rest.
Steep slopes still break navigation; that needs rough-terrain training of my own.
02 · Agents
Most of the robot work above was run by coding agents on a harness I built. I film the process and post it.
A local-first harness for long-running Claude Code and Codex work: 49 skills with evals, tickets, independent review and QA gates, scheduled automations. 830+ commits since March.
An agentic operating system for two Malaysian companies: it reads WhatsApp and meeting notes, keeps daily working memory and writes the weekly review back to the team's task board.
Build logs on Instagram: 23 Sep reel · 28 Sep reel
03 · Work
2026 – now
Robot learning on low-cost arms and agentic systems for companies. Selected engagements: 60x (AI for private equity), Prosus (AI personal assistant), Isara Labs (deep-research agents).
2025
Coding-harness evals and a Kubernetes sandbox system on SWE-Bench; autonomous app-development workflows on an in-house agent runtime.
2024 – 2025
First hire on a back-office AI team. Financial-statement review automation (about USD 5M ROI), bank-statement matching from 65% to 90%, 12 firmwide AI trainings.
2022 – 2023
Moved deployment tooling to AWS, cutting storage cost 50%.
2020 – 2024
Dissertation on human motion generation with diffusion models.
04 · Feed
Robot runs, build logs, hackathon wins and side projects. Keep scrolling and older posts load in.