Things that had to work

The defence from the patent, running as software: 84-segment spectral monitoring plus SLAM feature tracking. It runs a cheap frame-level check first and escalates to full spatial analysis only when something looks wrong. It uses no trained model, so there is nothing to keep up to date.

An adversarial testing rig for self-driving perception stacks under physical-layer light injection. It attacks object detection, lane recognition and depth estimation in turn and measures what each one does. It will take mitigation modules later, so the same harness covers both halves.

I wrote the electronic control and the desktop instrumentation for plant-inspired poroelastic soft robots at NiFELab — the rigs behind the mechano-sensing experiments, and behind the tactile-sensing paper accepted to Soft Robotics.

An encoder–decoder network that takes a shape a sensor only partly saw and predicts the missing trajectory. My undergraduate thesis, and where the real-time shape-completion paper started.

An arm I built by hand, with a switchable manual / semi-automatic control panel. I wanted glove control; the parts cost more than I had, so I built the panel instead.

An ESP-based car you can drive from the same room or from another continent, over the open internet.

It reads the tank level and starts and stops the pump at a threshold, so nobody has to stand there watching the tank.