I build AI systems that have to actually work — in regulated environments, at production scale, under real SLAs. “Mostly works” isn’t an acceptable answer when the system is making decisions about someone’s insurance claim or drug information. For the last two years at Deloitte AI Studio, I’ve been architecting GenAI systems for Life Sciences clients. A three-agent LangGraph pipeline that cuts legal claim review from 2-3 hours to 3 minutes. An on-prem SLM fine-tuned on 80M medical tokens. One of these is the subject of a patent I filed as lead inventor earlier this year. I also work on Deloitte’s innovation pipeline — prototyping systems that cross into manufacturing (safety and quality-control vision) and financial services (conversational advisors, fraud detection). The work sits at the intersection of classical computer vision, generative AI, and edge deployment, which is usually where the hard problems live. Before Deloitte, I was the founding ML engineer at Zocket, where I built the Creative Ad Generation Engine — a Stable Diffusion XL stack with IP-Adapters and a RAG-driven prompt augmentation layer. That engine powered Zocket’s core product through its growth. It’s also where I learned that in production, the model is maybe 30% of the problem; the other 70% is evaluation, observability, and failure modes nobody warned you about. I care about fine-tuning small models for narrow domains (frontier models are often too expensive and too leaky), agentic systems that fail safely, and retrieval that actually recovers recall on long messy documents. I think most “AI agents” in 2026 are still pipelines with aspirations, and the teams doing interesting work are the ones being honest about that. At Deloitte I run the AI team’s hiring rubric, and teach GenAI to non-engineers — designers, PMs, and leadership who want to use AI without becoming AI engineers.