Open Source
Public software, models, and datasets for deployable AI and scientific computing.
I publish research software, models, and datasets with the implementation, evaluation, deployment notes, and documentation needed to understand and reuse them.
DART
Real-time open-vocabulary detection using shared frontier model backbones, batched decoding, distillation, and TensorRT.
Valdrada
Real-time digital twin of New York City compiled from public records. It simulates traffic and pedestrians, labels every frame with segmentation, depth, and 3D boxes, and has a Python API modeled on CARLA.
G1 Patrol
Unitree G1 humanoid navigation through crowded sidewalks, crosswalks, and construction zones modeled on New York City streets, trained in Isaac Lab with teacher-student distillation to lidar and depth policies.
UrbanOmniDetect
Calibration-free monocular 3D detection across ego-vehicle, infrastructure, and aerial viewpoints.
Loom
Analytical neural computer that runs compiled C programs inside a looped transformer, with ONNX/WebGPU demos, FPGA verification, and regression tests.
bikeped
Real-time fisheye perception, safety alerts, simulator scenarios, and conformance testing for urban intersections.
UrbanAnonymizer
Face and license plate anonymization for street-level images and video, with offline tracking that keeps masks in place when detections drop out.
Boundless
Unreal Engine 5 simulator and datasets for photorealistic urban object detection, with automatic 2D and 3D bounding boxes.
GPTune
Early GPT-2 fine-tuning and sampling toolkit for adapting the 774M-parameter model on a single GPU.
FlyBrainLab
Natural language and ontology interfaces, OrientDB graph querying, GPU simulation, and interactive scientific visualization.