Inspiration
Meetings create commitments, but the work of turning “I’ll book it” into a correct calendar event is still manual, easy to forget, and hard to audit. We wanted to build an agent that closes that gap safely: it should act automatically when the intent is clear, but ask for help only when human judgment actually matters.
What it does
Executive Ghost receives completed-meeting transcripts via a Transcriptonic webhook (chrome extension), extracts explicit scheduling commitments, resolves known contacts and preferences, interprets dates in the user’s timezone, checks Google Calendar availability, applies autonomy policies, creates calendar events and Google Meet links, and verifies the result through a separate calendar read. When it cannot safely proceed, it opens a contextual exception with the source meeting, relevant transcript excerpt, detected commitment, an aclarative question for the user and exactly what will happen after the user responds.
How we built it
We built Executive Ghost with Codex and GPT-5.6 Sol at Ultra effort for the mayority of the session, then Terra for the tweaks, using TypeScript across the stack. The frontend is a React/Vinext dashboard; the agent backend is Fastify/Node.js; Supabase provides Auth, Postgres, RLS, encrypted integration credentials, and durable state; Google APIs power real Calendar availability and event creation; and the Chrome extension TranscripTonic provides the free automatic meeting transcript webhook made from Google Meet captions. The orchestration loop is explicit: Observe → Interpret → Plan → Policy → Act → Verify → Recover or Escalate → Report. We designed idempotency, retries, read-back verification, policy versioning, and audit evidence as first-class product behavior.
Challenges we ran into
The hardest part was making autonomy trustworthy. We had to handle incomplete transcripts, “You” speaker attribution, unknown attendees, timezone interpretation, all-day calendar conflicts, policy changes, API retries, and the possibility that an event is created but cannot immediately be verified. We addressed this with structured commitments, deterministic policy gates, one-time exceptions, idempotency keys, equivalent-event reconciliation, and a verification phase that never blindly creates a duplicate. We also had to research for a free easy to use transcript tool with webhooks integration.
Accomplishments that we're proud of
A real webhook-to-calendar agent flow. Google Calendar OAuth integration, availability checking, event creation, Meet links, and independent read-back verification. An exception UX that preserves meeting context and tells users exactly what their decision will do. Persistent contact memory: once a user confirms an external person’s name and email, Executive Ghost can recognize them in future meetings. A safe one-time policy override for timing constraints, without silently changing the user’s policy. A full demo lab with deterministic scenarios for success, ambiguity, recovery, duplicate prevention, verification failure, and policy conflicts.
What we learned
Reliable agents need more than a strong model. They need durable state, explicit authority boundaries, idempotency, verification, and UX that makes uncertainty understandable. The best autonomous behavior is often invisible; the best intervention is brief, contextual, and reversible.
What's next for Executive Ghost
We plan to add Outlook Calendar through the existing calendar-provider abstraction, richer contact resolution, follow-up workflows beyond scheduling, and integration with more transcript services like
Built With
- codex
- fastify
- google-calendar-api
- google-oauth
- gpt-5.6
- node.js
- openai-api
- postgresql
- react
- supabase
- transcriptonic
- typescript
- vinext
