Inkeep

The product

Inkeep is an agent platform with a no-code builder and a developer framework. It combines unified search and RAG over knowledge bases and docs with MCP integrations into a company’s apps and APIs. Its agents embed directly on a docs site to answer user questions, and they also review support tickets and product changes to draft doc updates.

The developer problem

An assistant on a docs site is only as good as what it retrieves. If a chunk splits a procedure in half or drops the version a page covers, the agent answers confidently from the wrong context, and users stop trusting it. Agents that draft doc updates depend on the same thing from the other direction: a draft is only right if the context behind it is current. So quality is decided below the model, in how content is chunked and labeled, how retrieval is measured, and how cleanly MCP tools expose a company’s real apps and APIs.

My focus

• Retrieval quality. I’d focus on measuring retrieval against a fixed set of real user questions, then tuning chunk boundaries and version metadata until the right passage comes back first.
• MCP tools and connectors. I’d design MCP tools that expose a company’s APIs as structured context, so agents answer from the real API surface instead of a plausible guess.
• Doc-update agents. I’d configure agents that turn support tickets and product changes into draft doc updates, with a human review gate before anything is published.
• Implementation guides. I’d write customer setup guides that I have run end to end, from connecting a knowledge source to embedding an agent on a docs site.

Why it fits my experience

• At Meta I built and maintain MCP servers in Python that expose SDK reference content as structured, retrievable context, so assistants answer from the real API surface rather than generating plausible method names that do not exist.
• At Meta I also restructured content specifically for machine retrieval: chunk boundaries placed at semantic units, explicit version and platform metadata attached to every chunk, and procedures written to remain correct when retrieved out of their surrounding context.
• At Red Hat I worked with the Support organization to identify which documentation gaps were generating tickets, and prioritized from that list rather than only from the feature launch schedule.