AI Chief of Staff

Autonomous executive assistant that monitors email, reasons about urgency, and only interrupts you on Telegram when it has to.

Comment

Inspiration

Professionals receive hundreds of emails daily. Most AI tools just forward everything — we wanted an agent that actually thinks and decides what matters.

What it does

AI Chief of Staff monitors your email through the Caspian SDK, runs each message through a 4-step LangGraph reasoning pipeline (classify intent → assess urgency → reason → act), and takes one of four actions:

  • Auto-reply for routine low-stakes messages
  • Send a Telegram approval card with Approve/Edit/Reject buttons for sensitive emails
  • Notify only when something needs attention but no reply is needed
  • Silently ignore spam and irrelevant messages

The agent learns from every approved reply and improves over time.

How we built it

  • Caspian SDK — one handler for Email + Telegram
  • LangGraph StateGraph — 8-node decision pipeline
  • Featherless.ai — OpenAI-compatible inference with Qwen2.5-7B
  • FastAPI — async API server with lifespan management
  • SQLAlchemy 2.0 — async ORM with SQLite (Postgres-ready)
  • Pydantic v2 — structured LLM outputs with validation and retry

Challenges we ran into

  • LangGraph drops state keys with underscore prefixes — learned to use unprefixed keys
  • Caspian SDK docs had conflicting information about method signatures — built compatibility fallbacks
  • Getting reliable structured JSON output from open-weight models required a retry loop with self-correction

Accomplishments we're proud of

  • One Caspian handler genuinely handles both channels with zero duplication
  • The approval workflow delivers replies back to the original email thread, not just Telegram
  • 18 passing tests with a deterministic fake LLM — no API spend in CI

What we learned

Building a real autonomous agent requires careful reasoning about when NOT to act. The hardest design decision was the approval gate — when should the agent ask permission vs act autonomously?

What's next

  • Calendar integration for scheduling awareness
  • Embeddings-based pattern matching for smarter auto-reply
  • Multi-user support

Built With

Share this project:

Updates

Submission history