I build AI systems that hold up for real users, every day.
Currently: GenAI in clinical settings, where "close enough" isn't good enough.
I got into this because I wanted to build things people actually use, not just prototypes that impress in a meeting. These days that means shipping both ends of a RAG system — the React interface someone clicks through, and the LangGraph orchestration deciding what it retrieves and why — with an obsessive focus on whether the answers can actually be trusted.
Every system I build has the same shape: a React client, a FastAPI layer, a vector store, and an LLM with guardrails. Here's what happens when a user asks NiDaan a clinical question.
Hover over a project to see it magnified. Each screen runs a live UI mockup of the actual system.
A retrieval-audited GenAI assistant for clinical queries. Users ask questions in natural language; the system grounds every answer in retrieved medical literature, cites sources, and flags uncertainty instead of hallucinating.
AI-assisted case analysis for legal teams. Paste a filing; the system extracts claims, surfaces precedent, and drafts annotated responses with citations. Built for speed and verifiable sourcing.
Interview prep platform with AI-driven feedback on spoken answers. Real-time transcription plus structured scoring across clarity, depth, and signal. Helps candidates iterate before the real thing.
A full-stack platform for discovering and streaming live radio stations worldwide through an interactive globe. Features an ML-powered recommendation engine, intelligent offline caching to improve reliability during API outages, and a FastAPI backend for personalized station discovery.
Not all skills are equal. The ones I reach for daily sit at the top — GenAI/ML is where I've been investing hardest, with full-stack as the foundation that lets me ship alone.
I didn't set out to work on retrieval pipelines and eval harnesses. I set out to build good interfaces. But building NiDaan — watching it hallucinate confidently, then fixing it one retrieval test at a time — changed what "good" meant to me.
Now the parts of the stack I care about are wherever trust gets built or lost: is the retrieval actually relevant, does the interface show its work, does the eval suite catch what a demo would hide. I move across the frontend, the backend, and the eval layer because that's genuinely where the interesting problems live.
Away from the screen: long runs, vinyl records, and a Neovim config in a permanent state of "almost right."
Open to AI engineering and full-stack roles, contract RAG work, and conversations about shipping GenAI to real users.