Ambitious AI.
Serious systems.

Build the execution layer that turns a model into a useful worker: its instructions, tools, reusable skills and boundaries. The harness makes the work repeatable and its outcomes inspectable.

Engineer the cycle of reasoning, action, observation and revision. Combine model judgment with explicit execution controls, and carry a multi-step task through interruptions and changing evidence.

Give agents distinct jobs and the infrastructure to work independently. Design delegation, routing and scaling around clear ownership, explicit contracts and isolated workspaces.

Build software-delivery loops that discover defects, implement repairs and bring a separately tested candidate to a release decision. Keep the original requirements and the actual tested revision connected throughout.

Make correctness observable. Evaluate model outputs and complete workflows, investigate failure paths, and require evidence for the behavior a system claims to deliver.

Connect answers to source material, then test the retrieval layer itself. Work across dense documents, vector and hybrid retrieval, and checks that distinguish relevant evidence from confident-looking output.

Build the pipeline between changing source material and an agent’s usable knowledge. Retain source revisions and exact passages, refresh embeddings incrementally, and keep retrieval tied to the reader’s access.

Build and integrate MCP tools, connect agents to external systems, and inspect what those connections actually return. Expose useful engineering checks as both a CLI and tools an agent can call.

Created Atlas at Knight CTO: an AI avatar with a conversational and knowledge system behind it. The engineering spans meeting media, speech, turn state, grounded answers and the lifecycle of a live session.

Built a Slack auto-responder that reasons over the conversation before deciding whether to reply. Connect event handling to fresh context, human handoff and a durable record of delivery.

Give long-running AI work a lifecycle beyond a single model call. Separate workflow control from stored state and evidence, retaining enough identity and ownership to recover deliberately.

Work on the model execution underneath the agent: routing, serving and local inference. Make provider selection explicit and investigate the latency, capacity and cost of the actual workflow.

Build the environment agents run in, with isolated execution, repeatable infrastructure and explicit deployment ownership. Bring enterprise distributed-systems experience to the operational work around AI.

Build around the customer’s data boundaries and the decisions people retain. Connect access controls and auditability to the workflow, including supervised AI in sensitive professional settings.

Take an idea through customer discovery, architecture, implementation and real use. At SettleOn, led a 10-person team and brought AI products to paying users; internal GTM agents supported the business behind them.

Stay close to the code while leading the decisions and people that make delivery possible. Combine daily engineering in Python, Rust and SQL with founder experience and engineering leadership at AWS and Workday.