projects

Equine biomechanics pipeline →

A multi-stage pipeline turning ordinary riding video into gait classification, joint-angle analysis, and per-frame residual signals. Starts at markerless pose estimation, ends at expert-validated scoring.

Trainer corpus suite →

Video-to-corpus generation: scene detection, parallel signal extraction (pose, action, emotion recognition), VLM synthesis, and cross-validated annotation. The failure analysis here is what defined REVEAL's problem space.

Dataset forensics →

Rebuilding a decaying deception-detection benchmark: a four-mode manifest-corruption taxonomy, a 96-99% identity-leakage audit of the official folds, audio-fingerprint re-alignment from broadcast sources, and verification gates designed to be able to fail.

vision_lab →

A React + FastAPI workbench for vision-language model interaction: multi-turn chat with prefix-cache-aware history, batch captioning with second-pass refinement, thinking-trace capture, and inline caption review.

github.com/aiaxmaior/vision_lab

Multimodal LoRA training harness →

A VLM fine-tuning harness for irregular hardware: asymmetric multi-GPU topology, memory-honest attention configuration, and the matched-control methodology used for attention-intervention studies.

github.com/aiaxmaior/lora-harness-showcase

AlphaZero-lite →

From-scratch AlphaZero-style chess engine: PUCT-guided MCTS with network-evaluated leaves and expert-iteration self-play on a single GPU. +241 Elo of self-improvement between generations, and a controlled AdamW-vs-Adam study worth +325 Elo with 95% CIs excluding zero. The search machinery here is the same family as the dispatch decision layer on /portfolio.

github.com/aiaxmaior/AlphaZero-lite

SORT multi-object tracker, from scratch →

SORT written from the filter up — a longhand 7-dimensional constant-velocity Kalman filter, IoU-cost association with Hungarian assignment, and a birth/coast/death lifecycle, in ~250 lines of numpy and scipy. Ships with ground-up MOT reference documentation, MOT16 I/O, a motmetrics evaluation harness, and a ByteTrack / BoT-SORT comparison runner. 34.2 pipeline FPS against ByteTrack's 33.5 on the same clip and detector; the tracker itself is ~1 ms/frame, so the number is detector-bound.

github.com/aiaxmaior/sort-mot-kit

TensorRT segmentation template →

A deployment template for lightweight real-time segmentation, with lane detection as the worked example: ONNX and TF-TRT build paths, FP16 and INT8, engine caching, dual-target Jetson/desktop deployment, and a classical OpenCV fallback behind the same interface when no engine is available. Includes the frame-budget arithmetic — 8 ms on an Orin leaves room for a detector inside 60 FPS.

github.com/aiaxmaior/trt-segmentation-template

Physics-driven steering model →

A zero-dependency steering model for driving simulators: lateral-acceleration envelope, bicycle-model understeer gradient, speed-scheduled first-order lag with slew limiting, and Ackermann per-wheel geometry. Written for CARLA, simulator-agnostic by construction. Every parameter is derived rather than tuned by feel — full derivations and a worked example from a vehicle spec sheet are in the repo.

github.com/aiaxmaior/physics-steering-model

Moza / Arduino HIL bridge for Linux →

Hardware-in-the-loop input and force feedback for simulators on Linux, where the vendor ecosystem offers neither: wheelbase and Arduino button-box events fused into one virtual gamepad over evdev/uinput, and telemetry-driven force feedback synthesized directly against the hid-universal-pidff kernel interface — spring, damper, road texture, collision impulse — with no game SDK in the path.

github.com/aiaxmaior/moza-linux-bridge

Edge perception systems →

The professional deployment record: real-time perception on Jetson-class devices, driver-safety simulation, and production recognition systems. Covered on the portfolio page.