Someone Built an AI Receptionist for a Mechanic Shop
This hit the front page of Hacker News today, and it's one of the most practical AI builds I've seen in months.
A developer built an AI phone agent for her brother's luxury mechanic shop. Not a demo. Not a proof of concept. An actual system handling real customer calls, quoting real prices, and collecting callbacks when it doesn't know the answer.
Here's why this matters for every indie hacker reading this.
The Problem She Solved
Her brother owns a mechanic shop. He's under the hood all day. Phone rings. He can't answer. Customer hangs up and calls someone else.
That's a lost job — sometimes $50, sometimes $2,000 — gone because no one picked up the phone.
She said he misses 100+ calls per week. Do the math on that. Even if 10% convert to actual jobs at $200 average, that's $2,000/week in lost revenue. $100K/year.
That's the kind of problem worth solving with AI. Not "summarize my emails." Not "write me a tweet." Real money on the table.
How She Built It (3-Part Breakdown)
Build the Brain (RAG Pipeline)
She scraped her brother's website — service pages, pricing, policies — and built a knowledge base of 21+ documents. Each document gets embedded as vectors (Voyage AI) and stored in MongoDB Atlas.
When a customer asks a question, the system retrieves the most relevant docs and passes them to Claude (sonnet-4-6) with a strict prompt: answer only from the knowledge base, keep it conversational, and if you don't know — say so.
No hallucination. No guessing prices. Just grounded answers.
Connect It to a Real Phone
She used Vapi as the voice platform. It handles phone numbers, speech-to-text (Deepgram), text-to-speech (ElevenLabs), and tool calling — all in one.
A FastAPI webhook server routes Vapi's tool-call requests through the RAG pipeline. She used Ngrok for local dev, would move to Railway or similar for production.
Every call gets logged to MongoDB: caller number, query, AI response, whether it escalated. The phone system becomes a data asset.
Tune for Voice
This is where most people stop early and ship something that sounds like a chatbot reading a webpage.
She tested 20+ AI voices from ElevenLabs. Found one called Christopher — calm, natural, sounds like someone who actually knows cars.
Rewrote the entire system prompt for voice: short sentences, no markdown, prices spoken naturally ("forty-five dollars" not "$45"), max 2-4 sentences per response.
The Exact Stack
Voice Platform
Vapi + Deepgram (STT) + ElevenLabs (TTS)
LLM
Anthropic Claude Sonnet 4.6
Embeddings
Voyage AI (voyage-3-large)
Database
MongoDB Atlas (vector search)
Backend
FastAPI + Uvicorn (Python)
Dev Tunnel
Ngrok
Why This Build Is Different
Most AI demo content online is useless for real businesses. It's either too abstract ("build an AI agent!") or too specific to the creator's workflow ("how I automated my newsletter").
This one is different because:
- It solves a money problem. Not a convenience problem. Missed calls = lost revenue. The ROI is obvious.
- The escalation path is a core feature. When the AI doesn't know something, it takes a callback. No dead ends.
- It grounds everything. RAG means no hallucinated prices. A raw LLM guessing brake costs would be worse than no AI at all.
- The voice matters more than the code. She spent more time picking the right voice and rewriting prompts for speech than building the backend.
What Indie Hackers Should Steal From This
1. Solve for money, not novelty. Every local business that misses calls is leaving cash on the table. Mechanics, dentists, salons, HVAC companies. The list is enormous.
2. RAG is the moat, not the LLM. Anyone can call the Claude API. The competitive advantage is a well-structured knowledge base that's specific to a business. That's your defensible product.
3. Voice is the next interface. Text chatbots are commoditized. Voice agents that actually sound human are where the value is moving. Get comfortable with Vapi, Bland, or Retell.
4. Ship the boring parts first. She built call logging, callback collection, and escalation flows before anything flashy. That's what makes it production-ready instead of a demo.
5. Test with real users, not yourself. Testing with your own questions in a terminal is step one. Step two is hearing how a confused customer actually interacts with it and fixing the gaps.
The Opportunity Is Massive
Think about how many small businesses still rely on a phone that someone physically answers. The plumber who's elbow-deep in a pipe. The dentist mid-procedure. The salon owner doing someone's hair.
AI voice agents can handle the first touch for all of them. And the stack is getting stupidly simple: Vapi for voice, a vector DB for knowledge, Claude or GPT for brains, and a cheap VPS to run it.
You don't need to build a SaaS. You can build one system for one business and charge $200-500/month. Ten clients = $2,000-5,000/month recurring. That's a real business.
The best AI products won't feel like AI. They'll feel like a really good employee who never calls in sick.
This Hacker News post is the blueprint. Go read the full breakdown if you want the technical details.
Original article: "How I Built an AI Receptionist for a Luxury Mechanic Shop"
Want to Build AI Agents That Actually Work?
The OpenClaw Ultimate Setup gives you the complete framework for building autonomous AI agents — the kind that handle your workflows while you sleep. Pre-configured, tested, and ready to deploy.
$29 — Get the Ultimate Setup