01
Agent systems that survive production
Specialized agents and skills on a real warehouse foundation. Memory structures that cut early hallucination. Bounded tools, not unbounded demos.
Agentic AISkillsRAGLangChain
Director of Data · Vitable HealthArizona
Warehouse · Agents · MCP
I build the layer agents stand on.
Seventeen years of warehouse and cloud work, then Head of AI at DVx, then founding Director of Data at Vitable. I still write the dbt and Terraform myself, and I still take the IC ticket when a team is blocked.
Control plane
Warehouse truth in, agent tools out. Click a node to navigate.
MCP → OpenAPI specs → typed MCP tool servers
0
Years in the stack
0
Portfolio companies advised
0
Industries shipped in
0M+
Lines of code
The journey
Seventeen years, one through-line.
From retail analytics in Flagstaff to founding data seats and a Head of AI tour — every stop shipped the warehouse, the pipelines, and the team. Drag the canvas and click any seat for the story.
My Journey
Full-time seatConsulting
Scroll to continue · click the timeline to explore
Practice
Pioneer in the stack, still on the ticket.
Agents, warehouses, and MCP are the same job for me: a foundation people can query, and tools an agent can trust.
02
MCP servers in Go
Custom MCP so agents call live APIs. Shared HTTP client, pooled connections, spec parsed once at startup. I wrote the open-source OpenAPI proxy that way.
MCPGoOpenAPI
03
Warehouses GTM actually uses
dbt models for renewal and expansion. Reverse ETL back into HubSpot. Finance and Ops query the same layer. Self-serve BI that left Metabase for Lightdash.
dbtBigQueryHightouchLightdash
04
Realtime pipelines and migrations
Postgres to BigQuery in realtime. AWS-to-GCP moves I led myself. Threat detection streams. Terraform for IAM, Glue, S3, and the path to ship.
PrefectAirflowTerraformKubernetes
Featured project
MCP OpenAPI Proxy
A Go MCP server that turns an OpenAPI spec into tools an agent can actually call.
Open source2025
Agents need live APIs, not screenshots of docs. Most OpenAPI-to-tool bridges are slow, parse the spec on every request, or hide auth behind a single happy path.
Write the proxy in Go. Share one HTTP client. Pool connections. Parse the spec once at startup. Expose every operation as a named tool with Bearer, API key, custom headers, and SSL options.
Why Go
- ▹Spec-driven tool generation. OpenAPI 3 specs parsed once at startup — $ref resolution and path-parameter handling included — so every operation becomes a named tool.
- ▹Agent-legible tool names. Names encode method and path (api_GET_users vs api_POST_admin_posts), with include/exclude regex filters and a 40-character cap for Cursor's combined tool limit.
- ▹One fast HTTP core. A single shared Go HTTP client with connection pooling keeps tool calls cheap under agent load.
Selected seats
From founding data to Head of AI.
Now
Looking for the next ambitious data and AI seat.
Principal or director-level work where the warehouse, the agents, and the team still need someone who will write the Terraform.