Domain 3 / 3
Embedded & Performance
Fifteen years shipping firmware under hard compute and memory constraints — the performance instinct that transfers straight to engine and tools work.
Gesture sensor platform with an on-device CNN
An automotive gesture sensor pitched to replace physical buttons across a vehicle interior — driven by a from-scratch, on-device neural network.
Problem
Replace physical controls (windows, hatch, lights, fuel flap, trunk) with a reliable gesture interface — classifying gestures in real time on a constrained automotive microcontroller (a Renesas RL78) with no room for a heavyweight ML runtime.
My Role
Defined the architecture and led development end to end: Flash API integration, sensor driver, LIN and UART stacks, data handlers, and the application module; gesture feature extraction (sample counts, peak timing, amplitude profiling across a multi-element optical sensor); a hand-written four-layer CNN in Embedded C fitted to the RL78's memory and cycle budget; and a C#/WPF companion app for recording gestures and managing training datasets.
Result
~98% gesture-classification accuracy at effectively instant on-device inference — end-to-end ownership from raw optical signal through the training pipeline to the classified gesture on the microcontroller. A rare, real 'embedded ML' credential.
Power-converter control on multi-core DSPs
AC/DC and DC/DC converter control on TI multi-core DSPs, MATLAB algorithms translated to the metal.
Problem
Drive AC/DC and DC/DC power-conversion hardware (including induction motor drivers) with tight, compute-intensive real-time control loops.
My Role
Defined the architecture and led development across all layers — base SW (PWM, ADC, SPI, CAN, SCI, IPC, QEP), middleware, and application — including translating MATLAB control algorithms into DSP code on TI's Control Law Accelerator, across multi-core controllers with inter-processor communication.
Result
Owned the full stack from silicon to control loop — including the MATLAB-to-silicon path most teams need two roles to cover.
Smart access door handle
Sensor drivers, algorithms, and comms for production smart car door handles.
Problem
Deliver reliable capacitive/inductive sensing and gesture detection in a smart access door handle, integrating work from multiple teams.
My Role
Architect and developer — responsible for the entire software implementation: sensor drivers and algorithms (capacitive, inductive), UART/LIN stacks, gesture detection, and base-SW integration of other teams' functionality.
Result
Architect-level ownership of all software on production smart-access door handles for a premium automotive OEM.