Patent-pending system that quantifies traumatic hemorrhage burden across three hemorrhage types and converts it into a clinical urgency score for intervention.

I built the labeled dataset from scratch: 3D masks proposed by a segmentation model, edited by the supervising radiologist and independently scored by a second, then linked to a clinical registry of interventions, transfusions, admission details, and outcomes. On top of it, the system combines 2D and 3D segmentation with slice- to case-level classification, per-case certainty quantification that makes the output explainable rather than a bare score, and a final urgency model trained with 50% feature dropout, which lets it run on imaging alone before admission labs arrive.

I own the evaluation gate as well: patient-level, out-of-time splits where every training case precedes every test case, per-class metrics rather than a flattering aggregate, and explicit threshold selection. Validated on a held-out temporal cohort of 3,480 trauma patients. Deployed at the R Adams Cowley Shock Trauma Center for prospective radiologist review and evaluation. Research use only; not FDA cleared. Patent pending.

PyTorch, nnU-Net, Vision Transformer, ConvNeXt, XGBoost, MC dropout, SimpleITK, DDP on 8xH100.

Fine-tuned Stanford's Merlin into the first 3D vision-language model for whole-body trauma CT, trained on 8,700 examinations with report-derived supervision and templated negatives and no manual annotation. Detects 83 distinct injury classes at a support-weighted mean AUC of 0.72, the broadest injury coverage of any published CT vision-language model. Accepted at SIIM-CAIMI 2026. 

Its labels come from a production LLM pipeline I built: 8,700+ free-text reports turned into sentence-level, entity-grounded labels over a 95-class ontology I designed, via a 37-rule extraction prompt, a rule-based negation engine and templated synthetic negatives, running on the OpenAI batch API at 50% lower cost, and it continues to label new studies as they arrive.

PyTorch, Merlin (3D ResNet + Clinical Longformer), report-derived supervision, mixed precision, OpenAI batch API, nibabel.

Rare findings do not have enough training data, so we generate it. SALIENT is a mask-conditioned diffusion model that synthesizes paired lesion-mask CT volumes: diffusion runs over discrete wavelet coefficients rather than pixels, under learnable frequency weighting, with a 3D VAE generating the masks and a semi-supervised EMA teacher pseudo-labeling the synthetic slices.

The result is synthetic data a detector can actually learn from, not just plausible-looking images. FID improves from 118.4 to 46.5, generation runs 28x faster than 3D pixel-space diffusion, and a downstream detector gains +0.12 AUPRC at 1% prevalence. We also characterized the augmentation dose-response: the best synthetic-to-real ratio rises as the labeled set shrinks. Preprint on arXiv; code release planned.

PyTorch, DDPM, discrete wavelet transform, 3D VAE, semi-supervised segmentation, ResNet-50, EViT.

Multimodal deep learning architecture for predicting wildfire ignition and spread from satellite imagery, weather, and grid telemetry, with a fault-prediction pipeline for power infrastructure. Published at CIGRE US and IEEE PES ISGT Europe.

PyTorch, CNN, Transformer, satellite imagery, geospatial raster data.

Deep CNN architecture for cooperative perception across two instrumented vehicles, sharing intermediate CNN features instead of raw point clouds to cut V2X communication overhead while improving detection accuracy. Trained and evaluated on CARLA synthetic point clouds while the real dataset was still being collected.

I specified and built the capture rig behind it, funded by Toyota: two Velodyne VLP-16 LiDARs, two GPS units and two 360 degree cameras. The hard part was not the models. It was synchronizing the LiDARs to microsecond accuracy over GPS PPS and NMEA, registering the two viewpoints spatially, and merging their point clouds in bird's-eye view so the same scene could be compared across sensors. I prototyped the pipeline on CARLA synthetic point clouds before any of the hardware existed.

TensorFlow, OpenCV, CNN, Velodyne VLP-16, VeloView, CARLA, point-cloud registration, BEV projection, sensor calibration and synchronization.