Clinical Intake -- AI Pre-Consultation Voice Agent

Architecture and implementation by Jinan Kordab. Copyright (c) 2026. All rights reserved.

Anonymous clinical interviewing pipeline with automated podcast generation.

---

video.mp4

Overview

A production-grade voice agent that automates pre-consultation clinical intake. The LLM conducts a structured interview by asking clinical questions aloud; the patient responds by speaking into a microphone. Every response is transcribed, analyzed, and synthesized into a structured audio podcast for physician review.

The system is designed for speed, privacy, and operational simplicity. A complete clinical questionnaire is processed in under three minutes from start to podcast delivery.

---

Architecture

Patient (Microphone)
      |
      | Web Audio Capture
      v
+------------------+      +------------------+      +------------------+
|   Groq STT       |----->|   LLM            |----->| ElevenLabs TTS   |
| Speech-to-Text   |      | Clinical Logic   |      | Text-to-Speech   |
+------------------+      +------------------+      +------------------+
                                    |                        |
                                    | Structured Data        | Audio Reply
                                    v                        v
                           +------------------+      Patient (Speaker)
                           |   SESSIONS/      |
                           | Podcast + Audio  |
                           +------------------+
                                    |
                                    | MP3 Download
                                    v
                           +------------------+
                           | Physician Review |
                           | (JWT Dashboard)  |
                           +------------------+

Pipeline flow: The LLM asks a clinical question via TTS. The patient answers by voice. Groq transcribes the answer. The LLM evaluates the response and asks the next relevant question. After all questions are answered, the full encounter is compiled into a downloadable MP3 podcast.

---

What Differentiates This System

Capability	Detail
Anonymous by Design	No patient identifiers stored. Audio is processed transiently; only the final structured podcast persists.
Sub-3-Minute Pipeline	End-to-end latency from first question to podcast delivery is kept under three minutes through model selection and pipeline optimization.
Full AI Stack	Groq for real-time STT, LLM for clinical reasoning and question routing, ElevenLabs for neural TTS with natural prosody.
Bilingual	French and English clinical questionnaires tailored to regional health authority requirements.
Auto-Generated Podcast	Each encounter produces a structured MP3 combining all questions and responses as a narrative summary for the physician.
Single-Command Deploy	docker compose up -d on any machine with Docker. No manual dependency installation required.
JWT-Secured Dashboard	Physician access to encounter recordings and session history behind token-based authentication.

---

Quick Start

git clone https://github.com/your-org/clinical-intake.git
cd clinical-intake

cp .env.example .env
# Add your API keys: GROQ_API_KEY, LLM_API_KEY, ELEVENLABS_API_KEY

docker compose up -d

Role	URL	Auth
Patient Intake	http://localhost:3000	None
Physician Login	http://localhost:3000/login.html	doctor / admin123

---

Environment Variables

Variable	Purpose	Provider
GROQ_API_KEY	Speech-to-Text transcription	groq.com
LLM_API_KEY	Clinical reasoning and question generation	Your LLM provider
ELEVENLABS_API_KEY	Text-to-Speech voice synthesis	elevenlabs.io
ADMIN_USER / ADMIN_PASSWORD	Physician dashboard credentials	Self-defined

---

Demo Recordings

The following MP3 files were generated by the system running inside Docker and downloaded from the container's SESSIONS/ volume. They represent complete clinical encounter podcasts as heard by the reviewing physician.

Recording	Language	Duration	File
French Clinical Encounter	French	~3 min	session-fr.mp3
English Clinical Encounter	English	~3 min	session-en.mp3

---

Test Suite

npm install
npm test

Tests spawn a real Express server on port 3099 using mock API credentials. No external service connectivity is required.

Suite	Tests	Result
Health Endpoint	1	Passing
Authentication (login, JWT, rejection)	6	Passing
Session Management (create, list, auth-gate)	4	Passing
Chat Pipeline (audio validation)	2	Passing
Podcast Generation (guard conditions)	1	Passing
Session Reset	1	Passing
Static Asset Serving	4	Passing
Total	19	100% Passing

---

Project Structure

.
├── Dockerfile                  # Multi-stage build, Node 22 Alpine
├── docker-compose.yml          # Single-command orchestration
├── entrypoint.sh               # Runtime permission bootstrap
├── server.js                   # Express API, AI pipeline, JWT auth
├── package.json                # Dependencies and scripts
├── jest.config.json            # Test runner configuration
├── .env.example                # Environment template (no secrets)
├── .dockerignore
├── public/
│   ├── index.html              # Patient intake interface
│   ├── login.html              # Physician authentication
│   └── doctor.html             # Physician dashboard
├── questions/
│   ├── fr.txt / fr_test.txt    # French clinical questionnaire
│   └── en.txt / en_test.txt    # English clinical questionnaire
├── img/                        # Brand assets
├── demo/                       # Sample podcast recordings
└── tests/
    └── server.test.js          # 19 integration tests

---

License

MIT. Copyright (c) 2026 Jinan Kordab. See LICENSE.