Featured · Research project
A brain-inspired operating system for drones.
Turning an ordinary drone with nothing but a cheap camera into something that would normally need thousands of dollars of sensors, so it can fly and make sense of a scene on its own in remote places. The vision pipeline mirrors the human visual hierarchy, every stage running concurrently on the GPU: a retina that wakes the heavy networks only when the view moves, occipital recognition for "what" and depth for "where", a parietal world model that keeps tracking smooth between detections, and a hippocampus that maps the space in 3D.
On top of it, a reinforcement-learning agent flies to targets on its own while avoiding obstacles, with a vision-language model coming next so it can take commands in plain language.
Brain-inspired architecture
Predictive-coding tracking
Reinforcement-learning navigation
Runs on any basic camera
Startup · Computer vision
CosineOne
Real-time AI security alerts for high-stakes sites. A full-stack build (FastAPI backend, JS frontend) with alerts surfaced both as a live overlay and through a Telegram bot. At its core is a private, on-device ML pipeline that validates custom natural-language alerts like "person taking something from the backyard." I built the pitch deck and demo, and engaged mining and defence prospects for pilots.
On-device MLLive + Telegram alerts
Startup · AI automation
TradieBuddy
An AI automation CRM that books jobs for tradies without a human in the loop. A scheduling algorithm uses the Google Maps API to slot new jobs into the calendar optimally, minimising travel between sites, and offers new leads the best available appointment instantly instead of waiting for an admin to call back. Faster bookings, fuller days, less drive time.
Route-optimised schedulingInstant lead booking
Contract · ML engineering
Semantic camera-roll search
Re-architected the backend for a semantic photo-search app, cutting average search latency from 180s to under 7s (about a 96% drop) while pushing retrieval accuracy to near-perfect. Built the text and image embedding pipelines, face and object tagging, and optimised inference to hit the production SLO. Also authored a reusable Python library for spinning up AI agents and wiring LLM tools into conversational photo search.
180s → <7s (≈96%)~100% retrieval