EdenicEdenic
← Selected work

AI · Voice · Alacrity

An AI receptionist that answers the phone.

Edenic engineers built Alacrity's multi-tenant AI voice receptionist — it picks up inbound calls, holds a natural conversation, captures the details, books the job, and texts a confirmation. A real-time bridge between the phone network and a live AI model, at sub-two-second latency.

PythonFastAPIOpenAI Realtime APIStreaming STTTwilio VoicePostgreSQLasyncpgCalendly APIAzure App ServiceAzure Key VaultDockerGitHub Actions
<2s

response latency

24/7

inbound call coverage

Multi-tenant

one platform, many businesses

Real-time

barge-in + live transcripts

Snapshot

Client

Alacrity Technologies — AI receptionist for US home-services businesses

What it does

Answers inbound calls, captures caller details, books the job, and confirms by text

Core

A live, bidirectional voice bridge between the phone network and a realtime AI model

Isolation

Multi-tenant by design — each business's data separated at every layer

The pain

  • Missed calls after hours and during busy periods meant lost jobs
  • Manual intake was slow, inconsistent, and easy to get wrong
  • No structured record of who called, why, or how urgent it was
  • Booking meant a callback later — friction that lost customers

Objectives

  • Answer every call, day or night, in a natural voice
  • Capture name, reason, address, and urgency as structured data
  • Book straight into the business calendar and confirm by SMS
  • Flag emergencies and keep every business's data strictly separate
System architecture — overview

Caller

Inbound call (PSTN)

Telephony

Twilio Voice
Media stream
Signature-verified webhooks

Realtime bridge

FastAPI service
Voice activity + barge-in
Booking + SMS

AI

Realtime model
Streaming transcription

Data

PostgreSQL
Transcripts
Appointments
Calendly booking + confirmation SMSPer-tenant isolation at every layerSecrets: Key VaultCI/CD: GitHub Actions → Azure

What we built

Real-time audio bridge

A concurrent bridge streams phone audio to a realtime AI model and back — no batch round-trips — keeping the whole speech-to-response loop under two seconds.

Natural barge-in

When the caller talks over the AI, the in-progress response is cancelled and queued audio cleared, so the conversation never talks over itself.

Streaming transcription

Speech is transcribed live with domain keyword biasing, and every utterance is saved in order for a perfect call record.

Booking + SMS

The AI books the appointment against the business calendar and sends an SMS confirmation, with delivery tracked end to end.

Per-business personality

Each tenant gets its own AI persona and intake rules without any code change.

Graceful goodbyes

Call-ending phrases end the call cleanly — with a guard that cancels if the caller starts speaking again.

Key takeaways

  • Every call answered — after-hours and overflow stop leaking revenue
  • Structured intake and live transcripts, not scribbled notes
  • Emergencies surfaced the moment they're mentioned
  • One multi-tenant platform onboards a new business with no code changes

Want to build something like this?

This is the calibre of engineer Edenic sources, vets, and places. Talk to a founder.

Talk to a founder ↗