VoiceBridge
A live Retell web call booked an appointment and completed calendar and CRM jobs through guarded releases.

Deployed · live Retell call booked a test appointment end to end
Independent build
My single-business voice-agent platform. A real Retell web call has booked an appointment end to end with a test caller; Vapi and phone-number acceptance remain open.
THE ENGINEERING
A deployed single-business voice-agent platform with owner sign-in, grounded knowledge, scoped Retell/Vapi and MCP tools, explicit booking confirmation and durable worker recovery. On October 25 an owner Retell web call joined audio, checked availability and booked a repair visit; the worker then completed the Google Calendar and HubSpot CRM jobs. The caller was a clearly labelled test voice; the agent, engine, worker, database, calendar and CRM were live. Four earlier calls exposed case-sensitive service matching, low caller audio, a missing year and an expired Google token, and each cause was fixed before the next call. Three guarded backend releases each took a database backup, kept a rollback target, matched 16/16 release files and 15/15 image modules and preserved 77 unrelated containers. The backend suite passed 258/258. The Vapi path, phone numbers, Google OAuth production publishing, a pre-call token check, provider-data deletion and backup replay remain open, and general follow-up email stays disabled. One booked call is a functional result, not a reliability or capacity measurement.
BEFORE → AFTER
Real problems from this build, what I changed and the verified result.
- CASE 01Before
A caller asked for “appliance repair” and the booking tool rejected it because service matching was case-sensitive.
AfterCall 5 ran availability and booking end to end; the Google Calendar and HubSpot CRM jobs completed for the booked visit.
How I fixed itMatched service names regardless of case and spacing, stored the configured name, and shipped it as a guarded release with a backup and rollback target.
- CASE 02Before
Availability failed mid-call because the Google refresh token expired after seven days while the OAuth app was in Testing.
AfterCalendar availability worked on the next call. Publishing the OAuth app and a pre-call token readiness check remain open.
How I fixed itRe-consented and rotated the token under a new explicit approval, released on the same image with 16/16 files and 15/15 modules matched.
- CASE 03Before
Owner downloads can expose another tenant's records or resurrect redacted transcript text.
AfterTwo live owner exports produced two durable audits without changing business records; nonempty, redacted and foreign-tenant cases were verified locally.
How I fixed itEnforced owner sessions, CSRF, tenant isolation, projected schemas, count/byte limits and durable redaction tombstones.
Evidence reviewed: Google Doc reviewed October 25, 2026; open gates stated above