Production document-intelligence pipeline - classification and field extraction over inbound clinical documents on Azure.
- Python
- Azure Document Intelligence
- Azure Functions
- Azure Blob Storage
- PostgreSQL
- Azure Key Vault
- Bicep
A few things worth reading about, then everything else. Each write-up covers what the problem was, what the approach turned out to be, and what actually came of it.
Production document-intelligence pipeline - classification and field extraction over inbound clinical documents on Azure.
An iOS cooking app on the App Store — scan a grocery receipt, get a pantry, and ask an assistant what to cook with it.
Shipped to the App Store
An agentic second brain - a Claude-operated vault, custom skills, and a dashboard that runs my week.
Semantic segmentation of seafloor imagery + bathymetry across 17 sediment classes, U-Net with ResNet backbones.
IoU ~0.74 on held-out set · 17 substrate classes
Operational analytics platform over a SQL reporting warehouse - utilization, billing, A/R and provider productivity.
A problem-mining agent — reads social platforms for real complaints, checks what already exists, and synthesizes grounded business hypotheses.
An agent-native hosting platform — deploy the small apps agents write, from a contract-first app spec to a running URL.
Successive optimization of matmul from naive loops through cache blocking to CUDA/cuBLAS, profiled with Nsight.
Evaluated a diffusion model for constrained protein backbone design - unconditional, symmetric oligomer and hotspot-guided binders.
Fine-tuned BERT for three-class tweet sentiment.
~92% accuracy on ~70k tweets
GRU recurrent network that predicts the winner of a chess game from the move sequence alone.
A fitness game where real workouts are the only way to grow your creatures — 28 species, no pay-to-win.
SwiftUI social fitness app - log workouts, visualize muscle-group activity, share progress.
Compared classical CRF against fine-tuned BERT for POS tagging and named-entity recognition.
Custom CNN vs fully-convolutional net vs fine-tuned VGG16 on CIFAR-100.
How detection models scale on the FathomNet underwater dataset - coverage, class balance and training feasibility.
Serial and parallel n-body simulation (OpenMP + MPI) and parallelized image convolution filters in C.