Maya: Voice AI Dental Receptionist

Autonomous, clinically-guardrailed Voice AI receptionist designed with Smile Dental Care, Bangalore (pilot in progress, integrating Hindi and Hinglish). Answers 24/7, verifies patient identities, schedules appointments against real-time SQLite Practice Management schedules, checks multi-payer insurance co-pays, dispatches WhatsApp confirmations, and safely triages dental emergencies with sub-second turnaround.

Launch Live Call Studio Live Observability Dashboard View GitHub Repo Back to Portfolio
100%
Urgent-case clinical recall
0.0%
Wrong-booking rate (read-back)
870-920 ms
Median conversational turnaround
28 / 28
Automated evals passing (100%)
Measured on 36 scripted test calls

1. Executive Summary & Clinical Context

Outpatient dental practices operate under severe front-desk constraints. Between greeting walk-in patients, sterilizing equipment, and processing billing, receptionists miss an average of 30% to 45% of inbound phone calls. In private dentistry, each missed new-patient call represents an immediate procedural loss of ₹3,500 to ₹15,000, and up to ₹50,000 in customer lifetime value.

Maya is a production-ready, clinically-hardened Voice AI agent engineered specifically for Smile Dental Care in Indiranagar, Bangalore. Maya operates over standard telephony and WebRTC browser audio, authenticating callers against real-time patient records, finding available provider slots, booking appointments, answering procedure-specific insurance co-pay inquiries across major Indian payers, and safely triaging dental emergencies.

Pilot Status & Multilingual Expansion: The pilot is in progress with Smile Dental Care, currently integrating Hindi and Hinglish conversational flows to support bilingual callers alongside Indian English.

Clinical-Grade Autonomy: Unlike casual open-ended chatbots, clinical voice agents must never hallucinate pricing, never prescribe treatments, and never miss an active emergency. Maya enforces deterministic validation, mandatory verbal read-backs, and conservative clinical escalation rules.

2. System Architecture & Tech Stack

The system is architected as an event-driven, low-latency pipeline connecting caller audio with a transactional Practice Management System (PMS):

Speech-To-Text (STT)

Deepgram Nova-2 with sub-200ms latency and custom acoustic boosting for Indian names, clinics, and locations (Aarav:5, Patil:5, Indiranagar:3).

Dialogue Engine (LLM)

OpenAI GPT-4o-mini with strict clinical negative guardrails, rapid time-to-first-token (TTFT ~250ms), and auditable JSON tool invocation.

Voice Synthesis (TTS)

ElevenLabs Turbo v2 streaming warm, natural conversational Indian English with sub-300ms synthesis latency and seamless interruption handling.

PMS Backend

FastAPI (Python 3.11+) async ASGI server providing tool dispatch, webhook ingestion, and live Swagger documentation.

Database Layer

SQLite 3 transactional storage with ACID compliance, relational foreign keys, provider schedule grids, and task escalation queues.

WhatsApp Gateway

1-Click Direct Links + Twilio API delivering instantaneous booking confirmations with clinic address and Google Maps links directly to WhatsApp.

3. Key Product Decisions & Latency Engineering

Decision 1: Mandatory Read-Back Before Booking (0.0% Error Rate)

Acoustic misunderstandings over telephone lines (e.g., "Tuesday at two" vs. "Thursday at two") lead to missed appointments, clinician downtime, and frustrated patients. Maya is strictly prohibited from executing book_appointment or reschedule_appointment without first reading back: Provider Name, Appointment Type, Date, and Time, and receiving an explicit verbal affirmative ("yes", "that works", "please confirm"). This resulted in a verified 0.0% wrong-booking rate across all 28 evaluation scenarios.

Decision 2: Sub-1.5s Conversational Turnaround Budget

Human conversational comfort breaks down when pauses exceed 1.5 seconds. Maya's budget was engineered for snappy, natural interactions:

Pipeline Component Observed Latency Optimization Mechanism
Deepgram STT 150 - 250 ms Nova-2 acoustic boost, endpointing tuned to 250ms silence
GPT-4o-mini TTFT 250 - 350 ms Concise system prompt, structured function schemas
FastAPI PMS Tools 30 - 80 ms Async SQLite queries, indexed slot & patient lookup
ElevenLabs TTS 200 - 300 ms Turbo v2 streaming audio chunks over WebSockets
Total Turnaround 870 - 920 ms (median) Comfortably sub-second; zero conversational awkwardness

Decision 3: Cognitive Load Limit (Max 2 to 3 Slots)

In telephone interactions, human auditory working memory is limited to $4 \pm 1$ items. Listing 5 or 6 open appointments over voice causes cognitive fatigue, forcing patients to ask receptionists to repeat themselves. Maya filters and reads back at most 2 to 3 curated slots (e.g. "Tomorrow morning at 10:00 AM with Priya, or Thursday afternoon at 2:00 PM with Dr. Rao"), reducing booking drop-offs and cutting average call duration by >40%.

Decision 4: Urgent-Case Recall as the North Star Metric

While generic SaaS chatbots optimize purely for Containment Rate, in healthcare operations, optimizing containment alone is hazardous. A system that books 99 cleanings but fails to triage a single acute dental abscess is unacceptable. Maya's primary North Star is 100% Urgent-Case Recall: if any symptom suggests an urgent or critical condition, Maya immediately allocates emergency slots or initiates staff escalation.

4. 11-Tool Practice Management Registry

Maya interacts with Smile Dental Care's PMS via 11 structured, auditable server-side tools:

Tool Name Parameters Operational Responsibility
find_patient full_name, dob Identifies existing patients. Enforces DOB verification to protect patient privacy.
get_available_slots appointment_type, preferred_date, time_of_day Queries open 2-week provider schedules (max 3 slots). Excludes reserved emergency holds.
book_appointment patient_id, slot_id, appointment_type Locks calendar slot atomically, prevents double-booking, and triggers WhatsApp confirmation.
create_patient full_name, dob, phone, insurance_plan Registers new caller into PMS with duplicate detection during live voice session.
find_appointments patient_id Retrieves upcoming appointments for authenticated caller; isolates cross-patient records.
reschedule_appointment appointment_id, new_slot_id Atomic slot swap: marks original booking 'rescheduled' and confirms new booking.
cancel_appointment appointment_id Releases slot back into open inventory and offers immediate rebooking options.
check_insurance_coverage plan_name, procedure_name Evaluates 15-row policy matrix across Star Health, HDFC ERGO, and ICICI Lombard for co-pays.
get_emergency_slot preferred_date? Unlocks Dr. Ananya Rao's reserved emergency buffer slots (12:30 PM & 5:30 PM).
create_callback_task patient_name, phone, reason Queues staff callback in tasks table for unhandled insurance, pricing, or custom requests.
transfer_to_human reason, patient_name?, phone? Initiates live receptionist handoff and logs audit event with type='handoff'.
send_booking_whatsapp patient_phone, appt_details Dispatches WhatsApp confirmation via Twilio API or generates 1-Click direct link.

5. Clinical Safety & Triage Protocols

Dental emergencies present distinct risks that must be triaged immediately:

Level 1, Same-Day Urgent Dental Care: Acute toothache, broken or knocked-out tooth, lost filling with pain. Protocol: Query get_emergency_slot to allocate Dr. Ananya Rao's protected emergency holds (12:30 PM & 5:30 PM). If filled, escalate via transfer_to_human with high priority.

Level 2, Immediate Medical Emergency: Difficulty breathing or swallowing, rapidly spreading facial swelling, uncontrolled oral bleeding, head trauma. Protocol: Maya instructs caller to proceed immediately to the nearest Emergency Room or dial 112. Maya explicitly refrains from booking clinic visits.

6. Evaluation Benchmark Suite (28 Scenarios, 100% Pass)

To ensure bulletproof reliability before production deployment, Maya was tested across 28 automated scenarios codified in pytest:

Category Scenario Description Scenarios Passed Pass Rate
Category A Happy-Path Routine Bookings (Cleanings, Fillings, Checkups, Ortho) 8 8 100%
Category B Insurance Inquiries (Star Health, HDFC ERGO, ICICI, Exclusions) 5 5 100%
Category C Clinical Emergency Triage (Level 1 Same-Day & Level 2 ER / 112) 6 6 100%
Category D Conversational Edge Cases (Interruption, Relative Dates, STT Spelling) 4 4 100%
Category E Out-of-Scope Requests (Billing Disputes, Medical Diagnoses, Complaints) 3 3 100%
Category F Patient Privacy & Authentication (DOB Mismatch Detection) 2 2 100%
TOTAL Full Codified Benchmark Suite 28 28 100% Pass

7. What This Demonstrates as an AI Product Manager

  • Voice System Architecture Ownership: Bridging low-latency WebRTC streaming, STT keyword boosting, and transactional database isolation.
  • Healthcare Governance & Risk Containment: Structuring clinical triage matrices that ensure zero missed medical emergencies and eliminate medical/financial hallucinations.
  • Auditory Behavioral UX Design: Engineering conversational protocols around human auditory chunking ($4 \pm 1$ items) and acoustic ambiguity.
  • End-to-End Observability: Codifying rigorous regression evaluation suites and building operational dashboards tracking containment, latency, and staff escalation queues.
Return to Portfolio