Systems designed to work beyond the demo.

End-to-end clinician monitoring system built in C# on Microsoft Azure (ML Studio, Data Factory) to classify sitting posture from sensor data, with anomaly detection flagging behavioural spikes and automated risk reports.

A multi-agent LLM system where a supervisor agent decomposes legal-document analysis tasks, routes them to specialist agents, and aggregates reviewed output — built and fully deployed in a single day in a domain I had no prior experience in.

Trained and INT8-quantised a YOLOv5 model, deployed to a LuckFox Pico RV1106 (256MB RAM, RISC-V, 1-TOPS NPU). Achieved sub-30ms on-device inference with zero cloud dependency, profiled under sustained load (37MB RAM, 65–68% CPU, 54.7°C thermal ceiling).

Designing and building a humanoid robot end to end: 6V high-current power architecture, I²C servo control across 19 actuators, firmware, and on-board vision and inference — managed independently from concept to hardware.

Register-level peripheral drivers written directly from the datasheet, bypassing the vendor HAL, with a systematic timing-analysis methodology documenting memory-mapped I/O behaviour.

Gas-optimised Solidity smart contract for transparent donation and withdrawal management, using a two-step propose/execute flow and AI-assisted evidence verification for withdrawal proposals. Built for a Web3 hackathon, deployed on Westend Asset Hub.

Android app with Firebase-backed login/registration, a product catalogue, and automated Excel quotation generation for a small sales workflow.