About

I'm a Lead AI Engineer at Tesana, where I co-architected the agentic engine that turns a text prompt into a playable 3D game. The engine was demoed live at GDC 2026.

I build agentic systems end to end: fine-tuning and evaluating models, optimizing inference to run on our own hardware, and designing the harnesses agents act inside. I also built real-time multimodal agents that play games from screenshots alone.

I'm self-taught. I started coding in 2023 while running a restaurant kitchen and was leading AI engineering less than three years later. What pulls me forward now is understanding how these models actually work inside: mechanistic interpretability, alignment, and AI security.

Experience

Projects

One of the first Google Calendar MCP servers, shipped months before Anthropic and OpenAI released built-in calendar tools. Referenced in arXiv 2603.23802 (UK AI Security Institute / Oxford).

A 3D character with real-time voice and long-term memory that lives across web, desktop, mobile, Telegram, and Minecraft as one continuous companion. Founder and sole engineer.

Research

I build harnesses that push frontier models as far as they'll go, then measure what they can really do. This is capability elicitation: a model that fails with a weak setup can often succeed with the right tools, scaffolding, and challenges, and good evaluations depend on finding that ceiling.

I run about 1,000 agents a month on hard, open-ended problems, with single runs lasting up to a day, and they keep doing things I didn't expect. I want to understand where those abilities come from, which is what draws me to evals, mechanistic interpretability, alignment, and AI security. Write-ups will land here as they're done.