AI Voice Agents — HelpDesk & Clinic Receptionist

Our Approach

This engagement delivered two AI voice agents on a shared platform: an IT helpdesk agent handling repetitive troubleshooting calls and a clinic receptionist agent managing appointment scheduling and caller routing. Both agents needed to operate in English and French with real-time performance.

Phase 1: Caller Identification & Voice Pipeline

The foundational challenge was caller identification. Traditional IVR systems rely on phone number lookup, but in a multi-site healthcare organization, employees call from personal devices, shared clinic phones, and different locations. We solved this with a novel approach: vector search on HR profile embeddings via Supabase pgvector.

When a caller speaks their name and department, the voice AI performs a real-time vector similarity search against embedded HR profiles (synced from Microsoft Graph), identifying the caller within seconds — no account number, no PIN, no IVR menu.

The voice pipeline was built on Twilio for telephony, Ultravox (fixie-ai/ultravox-70B) for speech-to-speech AI, and FastAPI for the orchestration layer. Language detection happens automatically in the first few seconds of the call, switching the agent's language model and response templates accordingly.

Phase 2: Helpdesk Agent — Troubleshooting & Ticket Creation

The IT helpdesk agent was designed for the highest-volume call categories:

• Password resets — guided self-service with identity verification
• Application access requests — automated provisioning via Microsoft Graph
• Common troubleshooting — RAG-based knowledge retrieval from IT runbooks
• Ticket creation — automatic logging with call transcript, intent classification, and priority assignment into Zendesk

Every action requiring a state change (ticket creation, password reset, access grant) includes an explicit confirmation step — the AI never takes action without the caller's verbal confirmation.

Phase 3: Clinic Receptionist Agent & Deployment

The clinic receptionist agent extends the same platform with a different knowledge base and tool set:

• Appointment booking — integration with the clinic scheduling system
• Caller routing — intelligent transfer to the appropriate department based on caller intent
• After-hours coverage — full capability outside business hours, replacing voicemail

Both agents share the same infrastructure, authentication layer, and monitoring dashboard, with knowledge bases and tool permissions isolated per agent type. The platform now handles calls across multiple sites with sub-3-second response latency and full bilingual support.