Physical AI · Perception · Real-time systems
I build perception and ML systems that connect research with working applications: grounding objects in 3D, evaluating robot policies, searching visual data, and making inference practical under latency and memory constraints.
Perception, learning and deployment
Independent implementations and adaptations spanning physical AI and ML engineering, alongside research and real-hardware work.
Perception · Completed
Combined pretrained detection and segmentation with depth filtering and calibrated camera-to-workspace transforms to locate named tabletop objects and estimate their dimensions.
Inference · Completed
Exported a PyTorch vision model to ONNX and served it through a containerised API. Benchmarked prediction agreement, memory and end-to-end latency under concurrent requests.
Data quality · Completed
Linked robot video with state and action traces to flag frozen frames, timestamp gaps and action discontinuities, then export reviewed training subsets.
Multimodal retrieval · Completed
Built natural-language search with visual embeddings, vector indexing, timestamped clip retrieval and duplicate suppression.
Robot learning · Completed
Reproduced and adapted an image-conditioned Diffusion Policy for PushT. Evaluated visual augmentation and action-execution horizons through controlled simulation rollouts.
Document retrieval · Completed
Built visual retrieval over engineering manuals and datasheets, returning source-linked pages for questions about diagrams and tables.
Research and robotics
Built and deployed PCB-defect classification and ADAS detection on a Xilinx ZCU104, including a custom Verilog 3×3 convolution accelerator, INT8 quantisation, and confidence recalibration.
Implemented voxelised LiDAR initialisation for Gaussian semantic occupancy and a mask-aware evaluator reporting per-class IoU, GPU latency and peak memory. I am designing confidence-gated temporal updates to reuse reliable geometry and prioritise uncertain or changed regions, with the aim of reducing recomputation while preserving accuracy. Research target: ICCV 2027.
Deployed and evaluated VLA and imitation-learning policies including π0.5, GR00T N1, ACT and Diffusion Policy across UR5/UR5e, Unitree G1 and RoboParty V1. Built ROS 2 and MoveIt 2 Servo integrations, RGB-D calibration and trajectory-auditing tools, plus Docker-based Isaac Sim/Isaac Lab workflows with a shared simulation and hardware control path.
Built a task-specific safety layer from recorded successful trajectories. It validates requested motion against the observed joint-angle and joint-speed space, rejects commands outside that envelope, terminates the run, and returns an explicit diagnostic. Next direction: predict commanded trajectories from current joint positions and velocities so the system can flag gripper-to-body impacts at damaging speeds before execution.
Technical depth + commercial context
My technical work is complemented by business-development and product experience, which shapes how I evaluate usefulness, adoption, and scale.
VLA and imitation-learning integration, PPO, motion retargeting, RGB-D perception, and simulation-to-hardware evaluation. Cross-validated Isaac PhysX and MuJoCo models against real torque measurements. Developing geometry-conditioned ACT for in-hand reorientation; manuscript in preparation for IROS 2027 with NTU MARS Lab.
Implemented LiDAR initialisation and occupancy-evaluation tooling under Prof. Tay Wee Peng, working with Dr. Loo Junn Yong. Designing confidence-gated temporal updates for efficient Gaussian occupancy; targeting ICCV 2027.
Industry partnerships and applied-AI initiatives with NVIDIA, Micron, and collaborators.
Led an AI analytics product, built event-to-price data pipelines, and analysed 20+ competitors to shape positioning and execution.
Deployment-aware edge-vision research with Dr. Loo Xi Sung, culminating in the IEEE ICIEA 2026 paper below.
Methods and engineering
Two application tracks
One-page resumes, updated October 2026.
IEEE ICIEA 2026 · Catania, Italy
Completed through URECA, the work covers classification, detection, custom hardware acceleration, quantisation, and confidence recalibration under measured performance and energy constraints.
Technical work with a reason to exist
I am interested in model capability, but also in what determines whether that capability becomes reliable, economical, and valuable in a real operation.
Better geometry and persistent state reduce ambiguity for planning, safety, and manipulation, which also reduces the cost of failures caused by incomplete scene understanding.
Batching, routing, selective computation, and memory-aware execution are what make larger models practical under real response-time and hardware constraints.
Robot utilisation, idle time, failure recovery, and the usefulness of generated data determine whether a technically capable system delivers value at scale.
Augment judgement, preserve accountability
I care a lot about using AI intelligently and responsibly. I use complementary agents, durable project context, explicit verification, and scheduled experiment monitoring because moving faster only matters if the work still holds up.
I use Codex and Claude for different parts of a problem, then get them to challenge and cross-check each other's work instead of trusting whichever answer arrives first.
I check claims against source files, logs, experiment results, and reproducible tests. If two agents disagree, I want that disagreement surfaced rather than quietly smoothed over.
I keep structured project records, handoffs, and status files so decisions don't disappear when a chat ends, which really matters when research runs across weeks or months.
I pair long-running RL training with scheduled agent check-ins that inspect progress, flag failures, and preserve results outside working hours, including weekends, while bringing decisions that need a person back to me.
Systems built and in development
This agenda combines an implemented runtime guardrail with active work on policy robustness, richer safety prediction, and the orchestration needed for large deployments.
Developing a simulation-first research direction around terrain, texture and lighting variation, with planned study of batching and serving under latency and compute constraints. Implementation baselines and performance results remain pending.
An implemented runtime layer derives an accepted joint-angle and joint-speed envelope from successful task trajectories, rejects motion outside that observed space, stops the run, and returns a clear diagnostic. The next direction predicts requested trajectories from the robot's current joint state to identify damaging gripper-to-body impacts before execution.
Foreman is my design for an orchestration layer that is installed into a site that already runs robots. It discovers and onboards them with operator approval, assigns missions across mixed fleets of robots and people, tracks progress, and recovers from operational exceptions. The plan and repository scaffold are done; robot integration is next.
