I build AI-native systems, agentic workflows, and infrastructure that holds up in production. My focus is on the gap between a model that can do something and a system that reliably does it.

• Spec-driven agentic pipelines — turning intent into implementable steps with enough structure to catch failures early
• MCP servers and context systems — practical interfaces between language models, tools, and real-world data
• Cloud and observability infrastructure — tracing, repeatable environments, and stacks that are debuggable under pressure

Models improvise when they lack context. I build around that: deterministic tooling, source-grounded retrieval, explicit interfaces, and failure modes that are visible rather than silent.

Most AI-assisted development breaks at the handoff between idea and implementation. Better specs, tighter feedback loops, and disciplined workflows make the handoff explicit instead of implicit.

Observability by default. Simple enough to reason about at 2am. I care more about tracing and repeatability than clever abstractions.

Building an AI infrastructure layer — the plumbing that makes agentic systems operable at scale:

• Federated MCP gateway — actions router with a thin IAM layer for agent registration and qualification (OAuth 2.1, OBO flows)
• Decision trace observability — capturing the full reasoning trail, not just inputs and outputs
• Lakehouse for agentic data — mining structured signal from SOR systems to ground agent decisions
• K8s-based orchestration — continuous tasks, long-horizon workflows, and on-demand workloads under one scheduler
• Local AI gateway — routing to cloud-hosted OSS models for cost and latency optimization

OSS vs managed is still an open decision.

• Clarity first. Speed follows from clear interfaces, not the other way around.
• Explicit tradeoffs over implicit assumptions.
• I work best where product ambiguity, technical depth, and execution pressure overlap — and someone needs to translate between them.