Work

Projects

Neptune answering a question in Claude Code: a collision stop on robot ARM-3A, with cited syslog, CMMS and incident records

Neptune (ongoing)

Creating a persistent memory system for a robot fleet; mapping all the context data and robot logs of a deployment environment into one live model.

Drone VLM System

Ask natural-language questions about live drone video. Qwen3-VL-2B base with a LoRA adapter, built as an extension of the aerial tracking system.

Aerial footage with detection boxes and track IDs

Drone Aerial Detection + Tracking

A YOLO26n detector fine-tuned on VisDrone, running on low-power onboard hardware (MaixCAM2 / AX630C). Switches between ByteTrack and BoT-SORT depending on whether the drone is hovering or moving.

First page of the World-Action Model survey

World Action Model Data Survey

Maps training data for world action models by causal completeness, and evaluates the data-collection market and the missing benchmarks for data quality. >200 downloads in less than a month of publication

ACT training loss and gradient-norm curves

Imitation Learning for SO-101 Arm

Technical exercise I did to learn about ACT imitation learning techniques. A ROS 2–MuJoCo bridge, an episode recorder and a LeRobot conversion path, then an ACT policy trained and run on the recorded demonstrations.

Hackathons

Aperture interface showing RGB video, hand landmarks, object detections and action segments

Aperture (1st place)

A phone-based marketplace where people record egocentric video for robot training data and get paid. Fuses video with LiDAR and IMU, with automated annotation.

Airtree Pathfinders

Uses AI to seamlessly analyse your photo library, on-device, to suggest events to go to with friends.

Polymates presentation slide at the UNSW DevSoc × Atlassian hackathon

Polymates (1st place)

First-of-its-kind social prediction market: bet virtual tokens on any event you want, climb your group’s leaderboard.

Build Club project dashboard showing client verification and risk-screening results

Build Club Hackathon (1st place)

An AI-assisted app that generates business risk reports for Tranche 2 compliance, cutting the process from weeks to a few hours.