I built an AI receptionist for a mechanic shop
A developer built a detailed AI voice receptionist for his brother's mechanic shop, aimed at capturing lost business from missed calls. The post dives into the technical stack, showcasing a practical application of RAG and LLMs for real-world problems. However, the Hacker News community debated the customer experience implications of AI in a 'luxury' service and questioned the architectural choices.
The Lowdown
The author developed 'Axle,' an AI-powered voice receptionist, for his brother's "luxury" mechanic shop, Dane's Motorsport. The primary goal was to prevent the loss of thousands of dollars monthly due to hundreds of missed calls while the mechanic is occupied. This isn't a generic chatbot; it's a custom-built agent designed to answer specific questions, provide pricing, and collect callback information.
- Building the Brain (RAG Pipeline): The core intelligence relies on a Retrieval-Augmented Generation (RAG) pipeline. The author scraped Dane's website for services, pricing, hours, and policies, embedding this knowledge into MongoDB Atlas using Voyage AI for semantic retrieval. Anthropic Claude (sonnet-4-6) then generates responses, strictly grounded in this knowledge base to prevent hallucinations and offer callbacks when information is unavailable.
- Connecting to a Real Phone Number: Vapi serves as the voice platform, handling phone numbers, speech-to-text (Deepgram), text-to-speech (ElevenLabs), and real-time function calling. A FastAPI webhook server processes Vapi requests, routes queries to the RAG pipeline, and sends responses back. Ngrok facilitated local development, and conversation memory was integrated to maintain context across turns.
- Tuning for Voice: Significant effort was spent optimizing the AI's spoken delivery. This involved selecting a natural-sounding voice (Christopher from ElevenLabs) and rewriting the system prompt for conversational, concise responses, avoiding markdown and filler phrases. An escalation flow was also developed to handle unanswerable questions by collecting caller details.
The project uses a stack including Vapi, Ngrok, FastAPI, MongoDB Atlas, Voyage AI, Anthropic Claude, and Python. Future plans include calendar integration for direct bookings, SMS notifications for callbacks, a management dashboard, and production deployment.
The Gossip
AI's Customer Conundrum
Many commenters expressed strong skepticism or outright dislike for interacting with an AI receptionist, especially for a 'luxury' service. They argued that AI undermines brand trust and customer experience, preferring human interaction, or even a voicemail, over a robot. A counter-argument suggested that an AI is better than no answer at all, particularly for small businesses where hiring a human receptionist might be unfeasible or too costly, and that modern AI might soon be indistinguishable from humans. There was also a debate about whether the "luxury" aspect of the mechanic shop was truly relevant or even accurate, with some suggesting the term was an AI-generated artifact in the post itself.
Architectural Analysis and Alternatives
The technical implementation sparked discussion regarding its necessity and efficiency. Several commenters questioned whether a full RAG pipeline and vector database were overkill for the relatively small, structured knowledge base of a mechanic shop, suggesting the information could likely fit within an LLM's context window. Others pointed out that LLMs, even with strict prompts, are still prone to hallucination, raising concerns about accuracy for critical business information like pricing. There was also a debate about why specific components (like Claude over GPT, or ElevenLabs for voice) were chosen.
Business Acumen vs. Technical Triumph
Beyond the technical merits, commenters questioned the core business problem and the proposed solution. Some suggested that a human outsourced receptionist would be a more straightforward and potentially cheaper solution, citing examples of services costing significantly less than the thousands lost from missed calls. Others hypothesized that if the mechanic is too busy to answer calls, he might already be at capacity, implying that taking on more work isn't the primary need. There was also a critical observation that the post lacked concrete production results or metrics on the AI's actual performance and customer acceptance.