muhammad.yousaf
All work
2025 – PresentTech Lead → Project ManagerClosed-source

Benjitron-9000

An AI agent that runs an executive's day across Trello, Google Workspace, and OpenPhone — every side-effecting action gated by a human approval before it fires.

FastAPILangGraphLangChainAzure OpenAIPostgreSQLSQLModel
3
Tool families
4
API endpoints
2–3
Team size

Summary

Agents that just *do* things break trust the moment they ship. Benjitron-9000 is built around LangGraph's interrupt-and-resume pattern: when the model picks a tool with a real-world side effect — a Trello write, a calendar invite, a text — the graph pauses on an approval gate until the user confirms. Wrong-send, wrong-card, wrong-invite all eliminated. Persistent PostgreSQL session state keeps the agent feeling like the same agent across days, not a chat app with amnesia.

Highlights

  • LangGraph state machine with interrupt + resume at every side-effect gate
  • SQLModel-backed memory store — sessions, messages, pending approvals, scratchpad
  • Three tool families wired with opinionated descriptions so the LLM never invents calls
This project is closed-source (built for a Kcube AI client). I'm happy to walk through the architecture, trade-offs, and code on request.