A luxury mechanic shop in Australia was missing over 100 calls a week. The owner, working under cars all day, couldn't answer the phone. Each missed call meant a lost job worth anywhere from $50 to $2,000.
So a developer in the family built Axle, an AI receptionist that picks up every call, answers service and pricing questions using real business data, and logs callbacks when it doesn't know the answer.
The stack is a good reference for anyone thinking about building voice AI for a small business. Vapi handles the telephony and call routing. Deepgram converts speech to text. ElevenLabs generates the spoken responses (the "Christopher" voice, specifically). On the backend, a FastAPI server processes each call through a RAG pipeline - meaning the AI doesn't just make things up. It searches a knowledge base of 21 documents covering services, pricing, hours, and shop policies, finds the relevant info, and passes it to Claude Sonnet as context. MongoDB Atlas with Voyage AI embeddings handles the semantic search.
The most interesting design decision: the system prompt tells Claude to only answer from the knowledge base. If a caller asks something not covered in those 21 documents, the AI doesn't guess. It collects the caller's info for a human callback instead. That's the right call for a business where quoting the wrong price on a transmission rebuild could be a $3,000 mistake.
Getting the voice right took iteration too. Written-sounding responses don't work when spoken aloud. The developer had to rewrite prompts for conversational delivery: shorter sentences, no markdown formatting, natural pauses.
The project isn't in production yet (deployment to Railway is pending), but the architecture is solid and the cost of the individual services is low enough to make sense for a single-location shop. For small business owners drowning in missed calls, this is a more practical template than most enterprise voice AI demos.