~/aicu labs
Hackathon builds, research demos and prototypes from the AICU team. Some become products, most don't. Every experiment comes with a write-up of what we tried and what we learned.
What are the HU values of air, lung, fat and muscle on CT? We measure them on a public chest CT with pydicom instead of copying a table, and show three traps real DICOM data sets: file order, -2048 HU padding and uint16 wrap-around.
How do you give each AI agent only the MCP tools it should have? We build a small gateway in front of the official git, fetch and time servers and measure it: tool tokens down 57-69% per role, about 0.6 ms per call, and 9 of 9 out-of-policy calls refused.
How do you build an MCP server, and what does it cost? We expose YOLO11 detection as MCP tools with the official Python SDK 2.x, then measure stdio overhead (0.6 ms), cold start (1.4 s) and why a base64 image as a tool argument costs 71,000+ tokens.
Which OSNet model should you use for person re-identification, and how fast is it? We load torchreid model-zoo weights, match YOLO11 person crops, export to ONNX and time four variants on a laptop CPU, with code you can run.
A hands-on look at using NVIDIA's Kimodo motion-diffusion model to synthesize rare gait patterns; and why the free per-frame foot-contact labels it hands back are the more interesting part for anyone training biomechanical models.
Running a detector on every camera frame in Flutter at 30 FPS: libyuv in C++ for YUV to RGB, native rotation and scaling, a GPU to NNAPI to CPU fallback on Android and Swift + Metal on iOS. First published June 2025, updated November 2026.
