Feedback and control
Getting a drone to go where you tell it is harder than it sounds. This is how the swarm does it.
Projects
Ferroflock
Sep 2024 – present
Sep 2024 – present
A drone swarm that flies like a flock, herded live in the lab.
Cheap drones are hard to fly precisely, and harder to fly together. The testbed is about ten DJI Tellos tracked by OptiTrack motion capture, and I wrote around 30,000 lines of C++ to run it. Each drone's position comes from motion capture blended with its own sensors, so it keeps flying when tracking drops out. My control loop sits around the drone's built-in one, and formation changes are checked for collisions before anything moves. The setup tools took lab setup from half a day to ten minutes.
Herding it shows why a flock model matters. Shepherd the drones from the edge and they stay together and move where you push them; walk into the middle and they scatter, which a rigid-body model of a flock can't capture. The second video shows it, with a tracked stick standing in for the herding drone.
It started as University of Sydney research and has been independent since December 2025, with Zihao Wang. The herding study is in the RA-L paper, whose drone-on-drone flight trials come next.
Coordination and flocking
Ten drones taking off without hitting each other, getting into formation, then flocking.
Setup tools and GUI
Setting up the swarm used to take half a day. Now it takes about ten minutes.
Tacita
May 2025 – Sep 2026
May 2025 – Sep 2026
Founder and AI systems architect. Blackbird Giants founder program, 2026.
Legal AI
An on-premise platform over more than 500,000 documents: case law, firm files and email. It delivered a 5,000-email eDiscovery matter end to end in under a week, from raw PST files through lawyer review to PDF bundles for US counsel, replacing months of manual work and A$30–40k in fees (the lawyers' estimate). At one point loading the database was projected to take months, because indexing had quietly slowed from 800 documents a minute to 20 or 30. After fixing the causes, a 200,000-document load takes an afternoon.
One lesson stuck. Its chronology fact-checker was marking every event as verified, and all 70 tests still passed, because none of them checked whether its verdicts were true. I replaced it with staged adversarial verification: the agents that do the work are separate from the checks that decide it's done, every claim needs an exact quote from the source, and a person signs off.
Substrate
An agent-native systems engineering platform for long-horizon work by people and agents together. A shared, typed system model links goals, requirements, architecture, decisions and verification, so less intent is lost at each handoff. Changes to canonical state go through reviewed proposals, and concurrent agent runs have durable approval gates and can resume.
Altus
A property acquisition and development platform that combines planning rules, parcel geometry, construction costs and market evidence into feasibility models. Its financial model runs monthly cashflows with correlated Monte Carlo over sale prices, build costs and approval times.
A drawing of an invented lot by the sea, with the street and the water to the east and a neighbour's house on each of the other three sides. A proposed three-storey building sits inside the setback line. Its shadows at 9 am, noon and 3 pm on 21 June fall to the south, and it hides part of the water from the house behind, uphill. The building can be moved with the arrow keys.
1Site plan
2Section A
Noon- Project
- Altus
- Site
- An invented lot
- Sun
- 21 June, 33.8° S
- Sheet
- Fig. 10
Fig. 10 — Three neighbours, one sea. You’re in the way. Argus and Episteme
Argus breaks investment theses into falsifiable subclaims tied to data feeds. Episteme generalises it into a belief engine for long-horizon work, where the facts keep changing: state estimation, but for beliefs. Each belief is a Beta distribution updated in log-odds, a source's reliability works like measurement noise and forgetting like process noise, independent sources count once, and every update is capped.
The night ferry is running again by June.
One claim's probability, drawn as the days pass, with its uncertainty band. Five sources of different reliability report yes or no. Each report moves the belief by a bounded step, a story another outlet runs again is counted once, the belief drifts back towards even odds as it forgets, and when the claim resolves the sources are graded. Choose a source to have it report.
Fig. 11 — One claim, five sources. Saying it twice doesn’t make it truer.
SLOSHSat
2024
2024
A CubeSat payload for watching water slosh in microgravity, the problem behind refuelling in orbit.
It was a group design project in my final-year space engineering course, taken through critical design review and prototyped on the bench. My parts were the custom circuit boards for sensing, localisation, experiment control and driving the magnetorquers, the firmware on FreeRTOS, and the structure, 3D-printed in carbon-fibre polycarbonate. Raw video of the tank would have been far too much to send down, so I wrote camera code to find the water on board and downlink only where it was.
A frame921,600 bytesSent down– 8 bytes
Earlier
Raytheon Australia
Automated engineering workflows on the Australian Defence Air Traffic System program, saving more than 500 hours a year across a five-person team.
Rio Tinto Centre for Mine Automation
Adaptive sampling on Pilbara drill-core data: 30% lower sampling density in sandbox experiments, and up to 50% more information per sample than uniform sampling.
University projects
A quadrotor that lands on a moving ground vehicle, and TurtleBot SLAM for disaster response.
Sydney University Rover Team
Founded and led a 25-person team building a Mars rover, through its critical design review.