Five small, focused systems — each built to demonstrate one thing well: signal processing, computer vision, decision-making under uncertainty, multimodal search, and systems-level deployment. Every number you see is measured, not scripted. Click into any one to see it running live.
A constant-velocity Kalman filter smoothing a live noisy signal in real time — the classical estimation-theory piece most ML portfolios skip.
Real-time object detection on your webcam or an uploaded photo, running entirely client-side — nothing ever leaves your device.
Thompson Sampling vs. a fixed 50/50 A/B split, racing live — watch cumulative regret diverge as the bandit learns which variant wins.
Search a photo gallery by meaning, not keywords — CLIP runs fully in your browser to rank images against any text you type.
Real, measured latency/throughput/model-size numbers across FP32 and INT8-quantized ONNX Runtime configurations — a systems-level validation report, not an estimate.
Every project has a Python reference implementation under its python/ folder — the "real"
engineering artifact, tested and in several cases run for real numbers you'll see quoted in the demo
(Kalman filter: 46% RMSE reduction on a synthetic trace; bandit lab: Thompson Sampling regret of 4.1 vs.
60.0 for fixed A/B; quantization benchmark: measured on this exact machine). The live browser demos under
each web/ folder reimplement the same algorithms client-side so you can interact with them
without a backend — full technical notes and honest caveats are in each project's own README.