I design agentic AI systems that survive production.

Architecture, data foundations, and operating practices for enterprises moving past the demo stage. This site is part portfolio, part digital garden — the thinking is public as it grows.

Work with me See the work Explore the garden

Reference builds — 5 published

Featured build LLM gateway as the control plane for agent systems One immutable policy snapshot controls model eligibility, privacy, residency, spend, routing, and typed fallback. Read the case study →

Data foundationsAgent memory over a lakehouse EvaluationEvaluation and observability harness GovernanceAgent privacy as a data-flow architecture Data operationsSutra Feed Guard

Coming next Durable agent execution with checkpoints and cancellation

Resumable runs, idempotent effects, explicit cancellation, and replayable failure evidence.

View all 5 builds →

AI systems — the focus

Most enterprise AI stalls between demo and production — because production AI is an architecture, data, and governance problem, not a model problem. That seam is where I work: agent architectures with properly designed memory, the data platforms that give agents something real to act on, and the evaluation and operating practices that keep them trustworthy.

• Agentic AI

  Agent architectures, memory design as a data modeling problem, evaluation harnesses, and failure-mode analysis for systems that act on our behalf.

• Data platforms for AI

  Lakehouse patterns, data modeling, and governance — the foundations that decide whether an agent system has anything true to work with.

Grounded in

The AI work stands on two decades across the full stack of practice: architecture & systems · product & strategy · UI/UX & design · platform engineering · teaching & training. The breadth is why the AI systems hold together; the garden holds the patterns from all of it.

Freshly tended

• Design tokens are the contract that makes AI-generated UI governable UI/UX
• Secrets discipline is a prerequisite for agent fleets, not an afterthought Platform
• Interpretability is a teaching problem — mechanistic insight only matters if humans can act on it Teaching
• Agent tenancy is a state-isolation problem before it is a compute-isolation problem Agentic AI

Now: building production-grade agentic AI reference systems and writing up the patterns as they prove out. What I'm focused on