Software developer Kedasha Kerr has published a detailed account of building Axle, a custom AI voice receptionist for her brother Dane's luxury mechanic shop. The system addresses a concrete pain point — missed inbound calls that Kerr says were costing the shop significant monthly revenue — by combining Vapi for telephony (Deepgram for speech-to-text, ElevenLabs for voice synthesis), Anthropic's Claude for response generation, and a <a href="/news/2026-03-14-captain-yc-w26-launches-automated-rag-platform-for-enterprise-ai-agents">retrieval-augmented generation pipeline</a> grounded in a 21-document knowledge base stored in MongoDB Atlas with Voyage AI embeddings. A FastAPI webhook server connects the voice layer to the RAG pipeline in real time. When the system cannot answer confidently, a fallback "saveCallback" tool captures caller details for a human follow-up.
The architecture reflects deliberate scope-limiting. Kerr constrained Claude to answer only from the curated knowledge base — a hard guardrail against hallucinated pricing, which would be particularly damaging in a trades context — and built the escalation path as a designed feature rather than something bolted on after the fact. Voice-specific prompt tuning required significant iteration: responses had to be rewritten to avoid markdown, bullet points, filler phrases, and numeric formatting that reads well on screen but sounds unnatural when spoken aloud. The build is documented as Part 1 of 3, covering RAG pipeline construction and basic telephony integration.
Hacker News commenters, including a former automotive service advisor, raised practical objections about production viability. The core issue is parts pricing volatility: any price Axle quotes can go stale within hours as supplier costs shift. One commenter pointed to California's Automotive Repair Act — enforced by the Bureau of Automotive Repair — and raised the question of whether verbal AI-generated quotes could fall under the statute's definition of a repair estimate, potentially creating liability when a final invoice differs from what the AI said. Whether similar exposure exists in other states is something Kerr and her brother would need to verify with a lawyer who knows their jurisdiction. The current Axle implementation includes no disclaimer that prices are subject to on-site inspection and a formal written estimate.
The practical problem is simpler than the legal framing makes it sound: a wrong price hurts the shop far more than it helps. A customer who gets a verbal quote of $400 and receives an invoice for $650 will dispute the charge; one who gets a callback instead of a quote will probably just book the appointment. That trade-off is where Kerr has already drawn the line with Axle — FAQ handling and appointment logistics on one side, dynamic pricing on the other. Two more installments are planned. The comments suggest the quoting question will be harder to solve than the wiring.