SeekOut MCP

The problem

An AI product can look convincing long before it is dependable. The harder work starts when it meets actual users, permissions, incomplete context, ambiguous requests, different assistant behaviors, and failures that do not repeat cleanly.

My role

I led the engineering work on SeekOut MCP from a customer-backed idea through public launch and stayed hands-on throughout. My scope crossed product shaping, architecture, implementation, evaluation, authentication, reliability, deployment, and the last-mile work required to publish the ChatGPT app.

I was promoted to Principal Software Engineer on October 1, 2026, and continue to work mainly as a hands-on individual contributor while managing a small AI engineering team. I collaborate with surrounding teams to get the work across the finish line.

The product decision

The product gives assistants a grounded way to work with recruiting search, profiles, talent-market questions, and recruiter context. The most important design goal was not making a polished demo. It was making the result understandable and useful when the request was messy.

Evolving with the protocol

As MCP evolved, I evolved the architecture with it. I introduced versioned, gateway-style routing and provider-specific endpoints that accommodate differences across AI clients. Shared authentication and operations stay consistent, while each client's authorization flow, available tools, and input schemas can account for its capabilities and integration quirks.

That work connects to the broader SeekOut Developer Platform: shared product capabilities, authentication, billing, and usage analytics that support developer integrations and AI-agent applications.

What I learned

Keep the team close to real user behavior, make quality measurable, and treat the awkward edges as product work rather than cleanup.