Research internship under the Defence Innovation Network program, developing sensor fusion and target tracking models.
Experience
An event recommendation and social coordination app for Sydney, live on the App Store and Google Play. I built and run the entire stack as the only engineer.
- Recommender. Personalized feed written in PL/pgSQL over pgvector. Each person is modelled as a handful of weighted interest centroids, updated as they use the app. The ranking spends some of its slots on safe matches and the rest on deliberate reaches. Written up here.
- Group scoring. Recommending to several people at once with a max-plus rule, so one strongly matched member can carry the choice instead of the group averaging into something nobody wanted.
- Data pipeline. An LLM ingest pipeline that scrapes unstructured listings and resolves them against what is already stored. It has turned about 90,000 raw events into 11,000 clean templates across 2,000 venues, with human review held under 1%.
- Serving and infrastructure. A Deno and TypeScript edge-function API covering search, the group itinerary builder, push notifications and place resolution. Row-level security on every table, instrumented with PostHog and Sentry.
- Frontend. The cross-platform app in React Native and Expo, on iOS, Android and a shareable webview.
Tutorials and lab sessions in robotics, state estimation and postgraduate AI.
- Robotics (MTRN4230). Forward and inverse kinematics, DH parameters and trajectory planning.
- State estimation (MTRN4010). Extended Kalman Filters and sensor fusion.
- Postgraduate AI (COMP9414). Search, machine learning and reasoning.
Joined as a summer intern and retained part-time through the academic year on the backend of Experion, Honeywell's flagship process-control platform.
- Developed backend features in modern C++ with Boost, and automated the nightly build-archiving process in Python.
- Worked in an Agile team, using JIRA, Confluence and Git.
Publications & research
Continuous Attribution on Discrete Spikes: Revealing the Explainability Paradox in Spiking Neural Networks
Neuromorphic Computing and Engineering (IOP) · first author
Class activation maps explain conventional networks well but break on spiking ones. The paper shows why, and proposes Poisson SmoothGrad, a noise model that keeps smoothed inputs within what an event sensor could produce. It benchmarks six attribution methods on three datasets and ends with a guide to choosing a method for each type of input.
Towards Closing the Domain Gap with Event Cameras
Australasian Conference on Robotics and Automation · first author
Compares event cameras with grayscale frames for end-to-end driving across day and night. Models trained on event data degrade much less under lighting shifts, since event cameras respond to relative brightness change rather than absolute intensity.
Calibrated Search Regions for Re-detecting Drones from Event Cameras
Preprint forthcoming · first author
When a tracker loses a drone, the detector has to search for it again. The paper compares calibrated search regions from classical motion filters and from a learned forecaster on event-camera data. At the same 90% coverage, the learned region is about five times smaller per window, and as the search window 0.8 s after a dropout it cuts a fixed box’s false positives by 37%.
Robotic Teleoperation with Haptic Feedback for Remote Ultrasounds
UNSW Sydney
A real-time haptic teleoperation interface between a Universal Robots UR5e arm and a 3D Systems Touch device over ROS and MoveIt, using quaternion-derived angular velocity, deadband and consecutive-zero filtering, and force feedback from the arm's own sensor.
Education
Thesis: Applications of Spiking Neural Networks in Robotics.
Supervised by Liao Wu and Francisco Cruz. Australian Government Research Training Program (RTP) Scholarship.
Honours thesis on robotic teleoperation with haptic feedback for remote ultrasound, supervised by Liao Wu.
Technical skills
Writing
How to build a recommender system in Postgres
The decisions behind a personalized feed on Postgres and pgvector: representing taste, choosing thresholds, candidate generation, popularity with few users, and cold start.
How to deduplicate scraped listings with an LLM
Normalizing records from a dozen sources that disagree, resolving them against what is already stored, and deciding which near-duplicates are the same event. Hango's ingest as the worked example.
How to keep a vision model working after dark
Why a model trained on event-camera data barely notices when the lights go out, what that costs elsewhere, and when the swap is worth making. From an ACRA 2025 paper on the day-night domain gap.