Speed-to-Lead AI Agent for US Real-Estate Brokerages

US real-estate leads go cold within minutes. Brokerages pay for leads from their website, Facebook and Zillow, then reply hours later, by which time the buyer has already spoken to another agent.

An AI inside sales agent that texts every new lead in under 60 seconds, day or night, qualifies them in a natural SMS conversation, books showings on the agent's real Google Calendar, and syncs everything to the brokerage's CRM (Follow Up Boss or GoHighLevel).

One intake for four lead sources: website forms, Facebook Lead Ads, Zillow and parsed lead emails, with de-duplication and phone normalisation.

Fast first touch: the opening text is template-built so the 60-second target never waits on a model. Claude takes over from the lead's first reply.

The model extracts, code decides: Claude returns structured facts, and a deterministic, unit-tested score marks the lead hot, warm or cold.

Compliance in code: no consent record means no text, STOP and revocation phrases are handled before any AI call, and quiet hours follow the lead's own time zone.

Human takeover: agents can jump into any thread, and a watchdog flags any lead that misses the 60-second target.

Proof for the client: a weekly digest shows response times, conversations and booked showings.

TypeScript on a self-hosted Windmill workspace (one small server runs many clients), Twilio for SMS, Claude for the conversation, and Postgres holding leads, messages, consent records and a full audit log. 91 unit tests cover the correctness-critical paths, plus an end-to-end test of the whole pipeline with external APIs mocked.