About
Engineering discipline for LLM applications and the systems underneath them
I build production LLM applications, multi-agent workflows, and the backend platforms they depend on — bringing seventeen years of engineering discipline to AI systems that need to keep working in real conditions, not just impress in demos.
Introduction
I build LLM applications, multi-agent workflows, and the backend platforms underneath them — applying the engineering discipline of seventeen years of payments, messaging, and platform modernization to AI systems that have to keep working past the demo. I'm strongest when the AI ambition is real, the production constraints are real, and the team needs someone who can hold both ends of that problem in one head.
Core values
- Clarify the real problem before scaling the solution
- Ground LLM systems in real data, real constraints, real failure modes
- Treat memory, context, and approval as design surfaces — not afterthoughts
- Prefer governed collaboration over autonomous agents that drift unattended
- Keep the system underneath the model honest enough to debug at 2 AM
- Name the tradeoffs out loud — quiet ones become someone else's incident
Professional narrative
I work on LLM applications, multi-agent workflows, and the backend platforms underneath them. Seventeen years of engineering — payments, messaging, platform modernization — sit under the AI work, which is the part of my profile most teams find rare: someone who ships LLM systems and is fluent in the production engineering they depend on.
I am most useful where AI ambition runs into production constraints. A workflow that demos well but drifts under real inputs. A multi-agent setup that needs governance before it needs more capability. A backend that has to keep working while AI features are bolted onto it. I am comfortable holding both ends of that problem in one head — the LLM half and the systems half — and naming the tradeoffs out loud.
How I work with AI is deliberate. Memory and context as design surfaces, not afterthoughts. Approval gates, structured handoffs between specialist agents, auditable batch protocols. Engineering judgment over engineering performance — the goal is systems that hold up after the team rotates, not diagrams that look elegant in a deck.
What I will not ship: autonomy theater, AI-as-magic framing, demos hardened into commitments they cannot meet. If the architecture is dishonest about what the model can and cannot do, no amount of polish saves it once it hits production conditions. I would rather name the gap and rebuild the honest version than smooth over a system that is going to embarrass everyone six weeks in.