Elicon
← ALL REFERENCE ENGAGEMENTS
REFERENCE ENGAGEMENT
Healthcare

Voice Intake for a Multi-Site Clinic Group

Answer every call, capture structured patient details, hand off to a person the moment anything is clinical.

  • Python
  • FastAPI
  • OpenAI Realtime
  • Twilio
  • Postgres

Reference engagements illustrate how Elicon would approach a project of this kind — the problem, the architecture and the decisions. They are examples, not accounts of past client work.

THE SITUATION

The problem.

A clinic group that loses patients to busy lines in the morning and voicemail at night. The engagement builds a real-time voice intake system that answers every call, fills a validated intake form as the caller speaks, and transfers to a person — with the form and transcript on screen — the moment anything looks clinical, urgent or unclear.

Phone-only intake creates morning queues, missed calls after hours, and inconsistent data that causes rework downstream.

  • ·A small front desk covers several sites, so peak-hour calls are abandoned and appointments lost.
  • ·Every intake collects the same fields slightly differently, pushing correction work into billing.

APPROACH

How we'd run it.

  1. 01Run a real-time voice pipeline with a structured intake form behind it.
  2. 02Transfer to a person on any edge case: urgency, unclear speech, unrecognised insurance.
  3. 03Review flagged conversations regularly and tune against them.
  4. 04Encrypt all patient data at rest and in transit, with access events streamed to the clinic's security monitoring.

ARCHITECTURE

What gets built.

Telephony bridge

Twilio media streams bridged to a realtime speech model over websockets, with a FastAPI service holding call state. Callers can interrupt at any time and the system stops talking.

Structured intake state machine

Behind the conversation sits a form, not a transcript. Each field has a validator and a confidence level; the model fills the form conversationally and the form refuses bad data.

Safety & handoff router

A parallel classifier watches every turn for urgency, clinical questions, distress and repeated misunderstanding. Any trigger warm-transfers to a person with the partial form and transcript already on screen.

Records integration

Completed intakes write into the practice management system with idempotent keys. Ambiguous patient matches queue for a person to confirm rather than creating duplicates.

DECISIONS

The trade-offs.

Form first, conversation second

A free-form assistant that produces a summary sounds impressive and creates cleanup. A validated form produces records billing can use.

TRADE-OFFConversations are slightly more directed than an open assistant.

Aggressive handoff thresholds

In a clinical context the costs are not symmetric. A needless transfer wastes a minute; a missed urgent caller is unacceptable.

TRADE-OFFAutomation is deliberately lower than the technology allows, and the threshold is agreed in writing during scoping.

Warm transfer with context, not blind transfer

Repeating everything to a person is the worst experience in phone automation.

TRADE-OFFDeeper integration with the phone system than a simple forward.

STACK

Typical stack.

BACKEND
PythonFastAPIPostgresRedis
VOICE
OpenAI RealtimeTwilio Media StreamsWebSocketsCustom endpointing
INTEGRATION
Practice management APIIdempotent record matching
COMPLIANCE
Envelope encryptionSIEM streamingRetention jobsRole-scoped access

Work of this shape usually runs as AI Product Launch →

SIMILAR PROBLEM?

Tell us what you're dealing with.

One conversation with an engineer is usually enough to tell whether this is a few weeks of work, a few months, or something you shouldn't build at all.

Talk to an engineer →