Where I’ve shipped.

I joined Tyler Sorensen’s lab to work on the WebGPU backend for ggml-org/llama.cpp, alongside PhD student Reese Levine. The goal: run modern LLMs on the GPU already in your machine. No native install. No WebAssembly hacks.

My contributions span shader compilation, kernel correctness, and dispatch limits. That’s the work that closes the gap between WebGPU as a spec and WebGPU as a viable inference target. Alongside the kernel work I built webgpu-bench — the cross-browser benchmark the contributor community uses to validate PRs against real hardware. The throughline: make the backend trustworthy enough that someone shipping a product can reach for it.

Theta Tech AI is a small applied-AI consultancy in medical imaging and signals. I joined as the team scaled up on QuantaFlo. QuantaFlo is an FDA-pathway peripheral arterial disease detector. I grew with the project across three production phases. First, the original training and inference pipeline from zero. Then a ground-up rewrite of the training stack for FDA-submission auditability. Then the serverless ensemble inference framework that now runs in production at Semler Scientific.

In parallel I worked across the company’s radiology-AI engagements with university hospitals. For Wisconsin I built a multi-GPU cloud training stack for 3D MRI segmentation. The stack shipped to Johnson & Johnson for tumor-volume tracking. For Case Western I built a lung-lesion similarity retrieval system at 92% recall. The U.S. Department of Veterans Affairs deployed it. For Emory I rebuilt a published radiology-AI codebase from scratch so other research groups could reproduce and extend the methods. Three different problem shapes. Three different clients. The same underlying job: get a real model to a clinician at scale.

TexNano is the medical-AI lab at Amrita Vishwa Vidyapeetham. I joined in 2021 as lead engineer on EUS_ML. EUS_ML is the data and training foundation for the lab’s endoscopic-ultrasound work. I stayed four years across the platforms that grew on top of it. The pattern I cared about most: build the infrastructure that lets clinicians use the models without touching the ML stack.

That shaped everything I shipped here. The data and inference pipeline for pancreatic cancer detection. The real-time web labellers used during live procedures rather than retrospective review. The VR pediatric heart segmentation model trained on CT scans of congenital-heart-disease patients. That model is in clinical use as of 2025, feeding a VR surgical-planning visualization. The lab’s research program drew ~$590K cumulative McGovern Institute funding (2023–2025) across this body of work.

A long thread in financial inclusion through The Mifos Initiative and Apache Fineract CN. The work is banking software for unbanked populations. I started as a GSoC 2019 contributor on Mifos Mobile CN. Delivered the full proposal. Migrated the codebase from Java to Kotlin. Shipped the app’s first complete release.

I stayed on as a mentor across three subsequent cycles. In 2020 I mentored the same mobile project and was invited to the GSoC Mentor Summit. In 2021 and 2022 I co-mentored the Mifos/Fineract ML credit scorecard line under the Apache Software Foundation umbrella. The 2021 work was federated-learning research on Fineract. The 2022 work was credit-risk and fraud-detection with Airflow + H2O.ai pipelines. The 2021 mentee returned in 2022 as my co-mentor.