I build AI systems that hold up when it matters.
Three years shipping production clinical AI — ambient scribes, RAG pipelines, fine-tuned LLMs, and the evaluation infrastructure that makes them trustworthy. Most of that work happened in healthcare, where a wrong answer isn't a UX bug — it's a patient safety event.
I grew up in Gujarat, India, surrounded by a family that ran a manufacturing business. The problem-solving was always around physical systems — how do you make this process faster, cheaper, more reliable? I think that's where I got the bias toward systems thinking and the reflex to measure everything.
I came to the US for an M.S. in Computer Science at Illinois Institute of Technology, graduating May 2026. The first real production system I touched was a DICOM reporting pipeline at Sarjen Systems — medical computer vision across 200+ hospitals. A 65% throughput improvement sounds like a line on a resume; what it felt like was learning that the gap between "code that works" and "code that works at scale in a regulated environment" is where most of the actual engineering lives.
That shaped everything after. At Leap of Faith Technologies I built QLoRA-fine-tuned LLaMA models for clinical decision support, an agentic triage system, and a RAG pipeline for visit-note-to-EHR matching. At the current stealth medtech startup in SF I'm building HIPAA-compliant AI systems — I'll share more in interviews, not in public.
The research side runs in parallel. I'm a co-author on HealthAdminBench at ICLR 2026 — a Stanford-led benchmark for Computer Use Agents in healthcare administration. The paper surfaces something important: CUAs are already performing administrative tasks in healthcare and nobody has agreed on how to evaluate them. We tried to fix that.
Evaluation before optimization
A model you can't measure is a model you can't improve. I spend more time on eval infrastructure than most engineers are comfortable with — it pays off every time.
Ship a tracer bullet first
Before building layers, I build one complete vertical slice: input → processing → output, end to end. It finds the integration bugs that only appear when layers meet.
The bug is almost never where you think it is
In clinical AI, especially. I've learned to distrust the obvious explanation and run the experiment.
Write the README before the code
If I can't explain what a system does in plain English before building it, I don't understand the requirements yet.
Boring tech is usually right
PostgreSQL over a new vector DB. A well-tuned prompt over a complex pipeline. The interesting constraint is the problem, not the tooling.
- Shipping HIPAA-compliant AI systems at a stealth medtech startup in SF
- Building EEAAO — an open-source dev companion that lives in VSCode and routes tasks to specialized open-source agents
- Building in public: 30-day arc documenting every decision on this site
- Open to founding-engineer and ML eng roles at applied AI startups, Series A–C
Production clinical AI systems. Details under NDA. Happy to discuss in interviews.
Fine-tuned LLaMA models for clinical decision support via QLoRA. Built an agentic triage system, a RAG pipeline for visit-note to EHR matching, and productionized LLM services on AWS + GCP.
Refactored a medical computer-vision service for DICOM reporting workflows across 200+ hospitals. Improved throughput by 65%.
Coursework in machine learning, distributed systems, and applied AI. Research in Computer Use Agents and clinical NLP. Co-authored HealthAdminBench (ICLR 2026 workshop).