Record of building an AI reception system for an auto repair shop.

Software developer Kedashya Kerr began building an AI-powered reception system after her brother, who runs a luxury car repair shop, was missing hundreds of phone calls a week because he couldn't answer them while working on cars. Kerr has compiled information on her blog about how she built an AI reception system that answers questions by referring to real data such as price lists and business hours, and then calls back for anything it doesn't know the answer to.

How I Built an AI Receptionist for a Luxury Mechanic Shop - Part 1
https://www.itsthatlady.dev/blog/building-an-ai-receptionist-for-my-brother/

According to Mr. Carr, his brother's luxury car repair shop often couldn't answer the phone while working on cars, resulting in hundreds of missed calls each week and several thousand dollars (approximately 300,000 to 800,000 yen) in lost opportunities each month. So Mr. Carr began developing 'Axle,' an AI reception system to handle calls to the shop. His goal was not to create a chatbot that answers anything, but rather an AI reception system specifically for the shop that could answer questions about business hours, fees, payment methods, cancellation rules, warranty details, availability of loaner cars, and the types of cars they service.

Mr. Carr's initial priority was to prevent the AI from arbitrarily guessing prices and service details. For example, if it were to quote a lower price for brake maintenance than the actual cost, it would contradict the explanation given during the customer's visit. Therefore, they adopted a system called ' RAG ' that generates answers by referencing external data.

Mr. Carr extracted service details and pricing information from his brother's factory website and created a knowledge base consisting of more than 21 documents. He then implemented a system where, each time an inquiry came in, he would search for the document that matched the question and use only the information contained in that document to answer it.

Mr. Carr used MongoDB Atlas as the storage location for the information referenced by the AI, and converted each document into a 1024-dimensional vector using the Voyage AI 's ' voyage-3-large ' model before registering it. A vector is a representation of the meaning of a text using a sequence of numbers that include decimals.

The relevant documents retrieved in this way are passed to Anthropic's ' Claude Sonnet 4.6 ,' which is then instructed to respond with the following conditions: 'It should only provide short answers using information from its knowledge base' and 'If it doesn't know the answer, it should say so and then pass on the message.' Mr. Carr notes that when he asks 'How much does an oil change cost?' on the terminal, it can now answer something like, 'Normally, oil is $45, synthetic oil is $75, and the work takes about 30 minutes.'

In the next step, Mr. Carr connected this system to actual phone numbers. He uses Vapi for the voice infrastructure, Deepgram for speech recognition to convert spoken content into text, and ElevenLabs for speech synthesis to read out the generated answers.

In practice, when a question comes in during a call, Vapi sends a query to a Webhook server created with FastAPI , the Webhook server generates an answer using the RAG mechanism, and finally Vapi reads that answer aloud. During development, the server was running on the developer's PC, but Ngrok , a service that allows a locally running server to be temporarily exposed to the outside world, was made accessible externally, allowing Vapi to connect.

This AI reception system not only has the function to answer questions, but also has a mechanism in place for when the AI cannot handle the situation. Specifically, the VAPI side has two functions: 'answerQuestion' used to answer questions, and 'saveCallback' to receive the name and a phone number for a return call. Call records, the content of the inquiry, how the AI answered, and whether a transfer to a human was necessary are stored in MongoDB. Furthermore, callback requests that arise from questions that could not be answered are recorded in a separate storage location. This will allow them to understand what kinds of questions are asked most often, when calls are concentrated, and how much human intervention is needed.

Mr. Carr spent particular time adjusting whether the text sounded natural when heard. Text that looks fine when read on a screen can sound unnatural when read aloud. For example, bullet points, set phrases like 'Certainly!', and notations like '$45.00' tend to sound unnatural when spoken.

Therefore, Carr rewrote the system prompts for voice use, including 'keeping responses to two to four sentences,' 'avoiding Markdown syntax,' and 'using natural, conversational language for amounts like '45 dollars'.' He also tried about 20 different voices before finally choosing a calm voice that suited a repair shop.

Mr. Carr emphasized the importance of 'how to handle content that the AI doesn't know,' and instead of guessing answers to questions not in the knowledge base, the system informs the caller that it cannot answer and then takes their name and a phone number for a return call. Furthermore, he said they have prepared integrated tests to handle cases such as fraudulent requests from VAPI, cases where vector searches cannot find documents that are close enough, and cases where the caller does not leave a phone number for a return call. Mr. Carr stated that 'for business-oriented voice AI, the process of handing over to a human is not an exceptional process, but a core function.'

According to Mr. Carr, as of the blog post published on March 20, 2026, this AI reception system is capable of answering questions and accepting callback requests. In the future, plans include adding calendar integration to allow reservations to be made during a call, a system to notify users via SMS when a callback request is received, and the development of a management dashboard. Furthermore, they plan to release the system with enhanced security so that it can be actually operated on the cloud service ' Railway ,' which allows apps and servers to run on the internet.