How We Built an AI Car Diagnostic for a Swedish Automotive Company
Carmy AB is a Swedish automotive tech startup building CarmyAI, a web app that acts as a virtual mechanic for everyday drivers. Users describe a symptom, paste a fault code, or upload a photo and get a clear diagnosis, cost estimate, and step-by-step repair guide. Their goal is to help drivers know what is wrong with their car before they spend a cent at a shop. To bring this vision to market, they needed a technical partner who could own the entire product build from the ground up.
Carmy came to us at founding stage with a clear vision but no in-house engineering team. The product concept was strong. The technical reality was not simple.
The core problem was specificity. Generic AI responses were a non-starter. The product had to pull real manufacturer repair data, match it to a user's exact vehicle, and handle multiple input types including symptom descriptions, OBD fault codes, and photos. Vague answers would kill trust instantly.
On top of that, the business needed to launch commercially ready, not as a demo. That meant building in parallel:
They needed one partner to own the full build.
We engineered an AI-powered mechanic assistant using Claude's tool-use capabilities, built on a Next.js full-stack architecture and deployed on Vercel with Supabase as the data layer.
Intelligent Vehicle Identification: We built a multi-country license plate lookup system that connects to manufacturer databases across 9 countries. Users enter their plate number, and the system automatically resolves make, model, year, and internal vehicle identifiers.
AI Agent with 24 Specialized Diagnostic Tools: At the core is a Claude-powered agent equipped with 24 automotive tools, from engine oil specs to diesel aftertreatment diagnostics. Claude autonomously decides which tools to invoke, chaining up to 10 tool calls per turn for complex multi-step diagnoses.
Transparent Streaming Diagnostics: We implemented a real-time Server-Sent Events streaming pipeline that displays exactly how the user's diagnosis was reached, building trust with DIY owners making safety-critical decisions.
Cost-efficient Prompt Caching: We enabled Anthropic's prompt caching on system prompts and conversation history, significantly reducing token costs on follow-up messages.
Token-based Usage Metering: We built a granular token-based billing system that logs usage per tool-calling leg, giving users fair value across Free, Plus, and Pro tiers with Stripe handling subscriptions.
We have helped over 200 businesses grow their value and improve how they work through better software.
