Hemut-MCP
AI Engineer. Hemut (YC P25), since April 2026.
TypeScript, React, Node.js, Model Context Protocol, Pinecone, PostgreSQL, Vitest
I work on the copilot inside Hemut's logistics software. Each part of a question gets routed to the tools it needs, and multi-tool runs wait for approval.
- Questions in the regression bank I built, across 20 routing and context categories
- 333
- Organizations in the Postgres estate its database tools run against, each scoped by row-level security
- 9
More than one answer
Dispatch questions often have two or three parts, and retrieving tools for the whole message could miss the second part. I extended the copilot with per-aspect semantic retrieval: each part of the question gets its own tool search, and the calls that answer them are coordinated in one plan.
Approval before running
When a plan needs several tools, execution waits for the user to approve it. The model decides what to call, and a person decides whether it runs.
A bank of 333 questions
I built a regression bank of 333 questions across 20 routing and context categories, so a change to retrieval or prompts gets checked against the same questions every time instead of whatever I happened to try that day.
Tenant scope
The database tools run in a 9-organization PostgreSQL environment. I moved their access onto transaction-scoped row-level security, so the organization comes from the session and the database enforces it, not the prompt.
The same conversation, in Command
I built the embedded Command experience around the actual copilot chat instead of maintaining a second renderer. Closing and reopening it keeps the conversation and the draft, and moving into the full app carries the session, organization, and pending clarification with it.
Next: Sense AI