I am an ML engineer working on GPU inference, model compilers and production computer vision. At Vetology I have deployed 1,000+ production models trained on 8M+ X-rays and processed more than 12 million images, cutting inference latency by 30%. On the side I build TensorRT engines in C++ with custom CUDA kernels, and I find and fix correctness bugs in the compilers and converters that ML models run through.
currently : Data Scientist @ Vetology, Chicago
focus : GPU inference, model compilers and converters, computer vision
also : medical imaging, model validation
background : MS Data Science, Stevens Institute of TechnologyBugs I found with differential testing (thousands of generated models compared before and after conversion or compilation), then fixed with tests.
| Project | Contribution | Status |
|---|---|---|
| NVIDIA TensorRT (ONNX GraphSurgeon) | #4940: constant folding produced invalid models when a tensor fed several Casts | under review |
| NVIDIA TensorRT (ONNX GraphSurgeon) | #4942: constant folding silently changed results (10 x 2.5 became 20) and broke comparisons; issue #4941 | under review |
| PyTorch torchvision | #9758: Python roi_align disagreed with the C++/CUDA kernels for boxes outside the feature map |
under review |
| Google LiteRT Torch | #1467, #1468, #1469: the NHWC layout pass sorted, repeated and indexed the wrong axes after a conv | under review |
| Apple coremltools | #2928: constant folding turned -inf attention masks into NaN |
under review |
| Project MONAI | MONAILabel #1912: made numpymaxflow optional on Python 3.13+ | merged |
Most of the repos below are clean, tested implementations built to be read and run. Where a project would normally lean on a huge pretrained model or external service, it uses a small stand in so the tests stay offline and honest. Nothing here claims a result a run did not produce.
|
Medical imaging chest-xray-classification: multi label chest X-ray classifier, real mean AUC 0.832, with a validation report and an FDA style change control plan medical-image-segmentation: UNet++ and DeepLabV3+ with a Dice plus BCE loss tumor-grading: a PathMNIST tissue classifier, real validation accuracy near 0.92 dicom-toolkit: DICOM windowing, Hounsfield rescaling, and normalization, tested against the pydicom samples |
Validation and governance model-monitoring: drift detection with PSI, the KS test, and chi square medical-uncertainty-estimation: Monte Carlo Dropout and conformal prediction with coverage guarantees medical-explainability: Grad-CAM, Grad-CAM++, and occlusion sensitivity side by side ci-cd-for-ml: a CI pipeline that trains, tests, and gates on a quality bar |
|
Self supervised and foundation models ssl-comparison: SimCLR pretraining compared by a shared linear probe mae-pretraining: a Masked Autoencoder from scratch medical-foundation-model: a small self supervised encoder pretrained on MedMNIST with linear probing ijepa-implementation: the core I-JEPA masking, predictor, and EMA target mechanics |
GPU inference and deployment tensorrt-optimization: TensorRT FP16 and INT8 benchmarks, a C++ TensorRT pipeline with custom CUDA preprocessing kernels, and Nsight profiles deepstream-video-analytics: DeepStream 9.1 multi stream detection and tracking, 1 to 16 streams at FP32, FP16 and INT8 kidney-tumor-ct-segmentation: MONAI 3D CT segmentation with TensorRT FP16 deployment evaluated at the validated Dice onnx-deployment: ONNX export with parity checks, served on Kubernetes and measured under load |
Medical and pathology
cancer-survival-prediction · survival-prediction-imaging · weakly-supervised-pathology · cell-segmentation · pathology-foundation-models · whole-slide-image-pipeline · stain-normalization · medical-image-augmentation · federated-learning-medical
Self supervised, multimodal, and foundation
contrastive-learning-methods · linear-probing-benchmark · dinov2-finetuning · clip-finetuning · image-text-retrieval · multimodal-embeddings · visual-rag · llava-medical-vqa · vlm-comparison
Generative models
diffusion-transformer · diffusion-math-walkthrough · controlnet-finetuning · dreambooth-finetuning · image-editing-with-diffusion · medical-image-synthesis · gan-progression
Detection, tracking, and video
open-vocabulary-detection · multi-object-tracking · sam2-finetuning · sam2-video-tracking · action-recognition · temporal-action-localization · video-anomaly-detection · video-object-tracking
3D, sensors, and remote sensing
lidar-object-detection · 3d-object-detection · 3d-scene-reconstruction · depth-estimation · occupancy-prediction · multi-sensor-fusion · optical-flow · pose-estimation-6dof · satellite-foundation-model
Efficient models and edge
model-compression · tensorrt-optimization · onnx-deployment · quantization-aware-training · knowledge-distillation · edge-vision-pipeline · real-time-segmentation
MLOps and infrastructure
ml-project-template · distributed-training · triton-inference-server · experiment-tracking · feature-store · data-versioning
Agentic
vision-agent · medical-diagnosis-agent · auto-image-annotator · image-search-engine · multi-agent-ml-experiments
Fundamentals and from scratch
transformer-from-scratch · diffusion-model-from-scratch · autograd-engine · yolo-benchmark · vision-transformer-from-scratch · backpropagation-from-scratch · positional-encoding-variants · loss-functions-explained · paper-implementations