How Autotelic works: an AI-run design-build system
Autotelic is a Denver design-build-automate contractor run on its own AI platform. This is the map: eight modules, five project types, and the services inside each one, with what every tool actually does.
The core
Autotelic runs on a small AI operating layer behind the scenes: a COO Orchestrator that coordinates eleven specialist desks and a rotating set of build desks. The desks each own one lane of the business — the phone line, the site chat, the client portal, labor dispatch, the supplier hub, organic growth, business listings, social publishing, uptime watching, and marketing direction. The build desks are a separate loop: each one picks up a single ticket, works it in its own copy of the codebase, and opens a pull request for a person to review before anything merges. None of this replaces a person’s judgment on money, compliance, or anything destructive — those decisions stay with a person.
The desks
- Voice — Owns the phone line — call intake, transcription, and call-intelligence summaries.
- Chloe — Owns the site's scoping chat and lead capture.
- Portal — Owns the client portal and the field-ops message flow.
- Labor — Owns crew labor intake and dispatch for finish work.
- Supply — Owns the trade and supplier hub.
- Growth — Owns organic growth and the lead pipeline from search.
- Listings — Owns business-listing accuracy across directories.
- Social — Owns social publishing to the company's own channels.
- Fleet Ops — Watches the platform's uptime and raises the alarm if something public goes down.
- Head of Marketing — Directs the marketing lanes and reviews content quality across the other desks.
- Build desks (VM 201) — The Kimi worktree loop — one ticket in, a branch and a pull request out.
Artifacts, reviewed
A build desk opened a fix for a sync bug. The reviewer's first verdict, 21:45: revise — the required test suite failed. The desk fixed the suite. Second verdict, 22:06: approve. Merged at 22:55 the same evening. Two words, twice, and a green check in between: that is the whole gate.
Issue #458 asked a desk to submit the agent-booking MCP server to a free directory. Desk 2 read the repo's rules first ("this repo owns the public discovery surfaces; this task must not change /estimate or /design"), added four files, ran its own verification script, and opened the pull request with the run log attached. Reviewed and merged the same evening.
The one-day clock
- 02:30 Trade knowledge writes (daily)
- 03:00 Citation checks (daily)
- 03:15 City code-facts harvest (daily)
- 04:00 Monday keyword-ranking sync (weekly)
- 04:00 Sunday content-packages plan (weekly)
- 04:15 Pending call summaries (daily)
- 04:30 Content packages build (daily)
- 04:45 BARna power recovery (daily)
- 05:30 Grounded content cycle (daily)
- 05:45 Content triage (daily)
- 06:00 Monday review monitor (weekly)
- 06:05 Content learning pass (daily)
- 06:30 Monday indexing promotion (weekly)
- 07:00 Search Console sync (daily)
- 07:00 Monday weekly scorecard (weekly)
- 07:15 Lead re-forward sweep (daily)
- 07:35 Expired-holds digest (daily)
- 17:00 Team-call recap (the COO digest) (daily)
- every 15 minutes Muni knowledge drain (interval)
- every 2 hours Build-desk PR pulse (interval)
- every 5 minutes Fleet watch (interval)
- every 5 minutes Blog scheduled publish (interval)