/teal-sea

production systems

ai agents that have answered a limo brokerage's phone since 2025: they quote, hold the card, and hand the booking to a person. A live operation, against real money.

01what it does

In service since 2025 for a limo brokerage in two US markets. The agents answer the line, quote from live rate logic, close, and place the card hold. A person reviews the booking before anything is captured. Voice, sales, payment and fulfillment run as one state machine, not four demos wired together.

It answers on the first ring, at three in the morning, in two markets at once. It does not get tired, and it does not quote a price it cannot honour.

02the constraint everything else follows from

A confident wrong answer is more expensive than a slow one.

That one line decides the architecture. An agent that invents availability, quotes a price that does not exist, or charges a card twice does damage at a speed no human matches. So the system is built to refuse before it improvises.

03what refusing looks like

04where the vendor stops

The voice is a vendor. The pricing, the availability and the booking write stay on my endpoints. A model can hold the conversation; it does not hold the money.

next.js / typescript / supabase / postgres / n8n / stripe / twilio / elevenlabs / ringcentral / redis

05one truth, three layers

Humans and agents read the same truth and the same permissions. Neither carries a private copy of reality.

06the unglamorous half

A visibility-aware polling change removed roughly 23,000 requests a day from idle browser tabs.

Explicit database deadlines. Fail-closed settings. Cross-channel state. Observability that records what an agent actually did, not what it reported doing.

07built from inside the operation

I ran the operation before I wrote the software that runs it: operations manager from 2019, director of operations from 2020, one US market to eight. The constraints came off the dispatch floor, not off a whiteboard.

live since 2025 · a limo brokerage · two US markets · real bookings and card holds · running the operation since 2019

The model can reason.
It does not get to be its own authority.