Projects
Things I’ve built across computer vision, robotics, and autonomous driving — measured on held-out data where it applies, and written up honestly. Wins and dead ends alike.
Autonomy/Stack
liveAutonomous driving · 3D visualization
An interactive 3D walkthrough of a self-driving car's perception-to-control stack — sensors, compute, and the deep-learning models, stage by stage, on real nuScenes data. v1 covers Tesla FSD (camera-only, end-to-end), with primary-sourced claims and a failure gallery.
LenaLab
Computer vision · visual odometry
A Claude agent authors visual-odometry algorithms from scratch in a sealed sandbox; the verifier grades them on held-out trajectories. Best result: 0.033 m metric on an unseen scene, beating the classical reference.
Composite
Image processing · super-resolution
Multi-frame super-resolution from shaky dashcam frames: sub-pixel register 31 low-res shots, then composite them into one sharp grid that reads a plate no single frame can. The lever is a 60-year-old derivative-free optimizer (Nelder–Mead simplex) made to work on images by smoothing the cost with a Gaussian scale-space — checked on 50 sequences against pseudo-GT (NCC, no-norm: +0.27 dB mean, 64% win-rate).