Case Study · AI Systems

Reception System

An AI-powered phone handler for small businesses. Answers calls, extracts intent, books appointments — no internet, no cloud fees, no subscription.

Python FastAPI Faster Whisper LLaMA / Phi-3 Google Calendar API Ollama UWE Enterprise Scholarship

£1k

Enterprise scholarship

100%

Local inference

Pilot

Live deployment

The Problem

Small clinics and SMEs miss calls constantly — receptionists are expensive, call-centre services are impersonal, and most AI phone tools require cloud subscriptions that don't suit businesses handling sensitive data. A dental clinic can't route patient calls through a third-party cloud service without GDPR implications.

The ask was simple: answer calls, extract the reason, and book the appointment — all on hardware the business already owns, with no data leaving the building.

Architecture

Inbound call

→

SIP / VoIP adapter captures audio stream

Faster Whisper

→

CPU-optimised Whisper tiny model. Real-time STT, ~400ms latency on i5

Local LLM

→

LLaMA 3.1 / Phi-3 Mini via Ollama. Extracts intent, name, preferred time

FastAPI core

→

Orchestrates STT → LLM → Calendar pipeline. Async, typed, testable

Google Calendar

→

Checks availability and books slot via OAuth2 service account

Confirmation TTS

→

Text-to-speech reads back the booked slot to the caller

Key Challenges

→ Latency on commodity hardware. Whisper tiny runs at ~400ms transcription lag on a standard i5. Streaming the LLM response reduces perceived wait time significantly — the caller hears "I'm checking availability..." while the model is still processing.
→ Hallucination-safe slot booking. LLMs will confidently invent appointment times. The system validates all structured output against the live calendar before confirming anything — the LLM proposes, the calendar confirms.
→ Accents and background noise. Whisper handles accents well, but noisy environments (clinics are loud) require pre-processing. Added a noise gate and voice activity detection before the STT step.
→ Graceful failure. When the system can't extract a confident intent, it falls back to taking a message and sending a callback request rather than failing silently or hanging up.

Outcome

Currently in pilot deployment with a real client through the UWE Enterprise Scholarship (£1,000 award). The system handles inbound calls, books appointments, and sends confirmations without any human in the loop. Missed-call rate has dropped to near zero during pilot hours.

The project is intentionally not open-sourced yet — the pilot client has it running on their hardware. Architecture documentation and a demo version will be released after the pilot concludes.

← back to portfolio next: dissertation →