MechAI

slide deck

https://www.loom.com/share/49b60f0fa78e410bafee95f8bc4ae560 video link

Inspiration Every year, millions of drivers are stranded on the side of the road with no idea what's wrong with their car. They call a mechanic, struggle to describe the problem, wait hours for someone to show up, and often get misdiagnosed because the mechanic had no context before arriving. We asked: what if the moment a customer calls, an AI agent already knows exactly what's wrong, has sourced the parts, and has briefed the mechanic — before they even leave the shop? That question became MechAI.

What it does MechAI is a real-time AI-powered automotive service management dashboard. When a customer calls:

VAPI handles the voice call — an AI agent collects the customer's name, vehicle details, and describes symptoms The moment the call ends, an AI agent pipeline automatically fires: Creates a job card in the dashboard Runs GPT-4o to generate a full mechanic-grade diagnosis with root cause, severity, diagnostic steps, and parts list Runs Tavily to search RockAuto, AutoZone, O'Reilly and others for real part prices in real time Runs Modulate ToxMod to flag any abusive or toxic language in the call transcript Queries Neo4j for historical repair patterns on the same make/model Sends the customer a WhatsApp message (via Twilio) asking them to send photos of the issue When the customer replies with photos, Reka Vision analyzes the images and adds a visual inspection report to the job card The mechanic arrives briefed — with diagnosis, parts list, prices, and photo analysis already done How we built it Layer Technology Frontend React 18 + TypeScript + Vite + Tailwind CSS + shadcn/ui Backend Supabase (PostgreSQL + Edge Functions + Realtime + Storage) Voice AI VAPI — conversational voice agent AI Diagnosis OpenAI GPT-4o with structured tool calling Parts Search Tavily Search API → GPT-4o-mini for structured extraction Knowledge Graph Neo4j Aura — vehicle/symptom/diagnosis graph Voice Safety Modulate ToxMod API Visual Inspection Reka Vision (multimodal) Customer Photos Twilio WhatsApp → Supabase Storage → Reka Vision Agent architecture: Each Supabase Edge Function acts as a focused agent. The vapi-webhook agent orchestrates everything — it spawns the ai-diagnosis agent, search-parts agent, Modulate check, and WhatsApp notification concurrently in a fire-and-forget pattern so the mechanic's dashboard updates in real time without blocking.

Challenges we ran into Reka Vision URL fetching — Reka fetches images directly from their servers, so images must be at fully public HTTPS URLs. Private or bot-blocked URLs (like Wikipedia) silently fail. Solved by routing all photos through Supabase Storage with a public bucket. VAPI webhook auth — Supabase Edge Functions require JWT auth by default, but VAPI (an external service) sends unauthenticated POST requests. Functions had to be redeployed with --no-verify-jwt. Neo4j HTTP API vs Bolt — Neo4j Aura uses the neo4j+s:// protocol (Bolt), but Deno doesn't support Bolt drivers. Had to convert to Neo4j's HTTP transaction API by rewriting the URI scheme programmatically. Dead Supabase project — The original project was paused for 90+ days and is unrecoverable. Had to spin up a brand new Supabase project via the Management API and migrate everything Accomplishments that we're proud of Full agentic pipeline in under 60 seconds — from a customer hanging up the phone to a mechanic having a complete diagnosis, parts list, pricing, toxicity flag, and visual inspection report on their screen Self-improving knowledge graph — every call that goes through MechAI stores a Vehicle→Symptom→Diagnosis→Part pattern in Neo4j. Future calls on the same make/model get enriched with real historical data, increasing diagnostic confidence over time Zero mechanic input required — the entire intake, triage, diagnosis, parts sourcing, and photo analysis pipeline runs without a human touching anything WhatsApp → Reka Vision pipeline — customers can text photos from the roadside and have them AI-analyzed and attached to their job card automatically What we learned Multimodal agents compound value — combining voice (VAPI), vision (Reka), search (Tavily), and graph memory (Neo4j) produces results no single model could. Each agent handles what it's best at. Graph databases shine for domain knowledge — Neo4j's property graph model is a natural fit for automotive repair patterns. The Vehicle→Symptom→Diagnosis→Part structure maps exactly to how mechanics actually think. Toxicity detection needs to be passive — Modulate works best as a silent background check, not a gatekeeper. Flag and inform, don't block — the mechanic still needs the job even if the customer was rude. Public URLs are a first-class concern in multimodal AI — any vision API that fetches images server-side requires images to be truly public. Storage architecture decisions have to account for this from day one. What's next for MechAI SMS/MMS photo intake — extend the WhatsApp flow to standard MMS so any customer can send photos without WhatsApp Mechanic mobile app — real-time push notifications when a job is assigned, with offline-capable diagnosis view Parts ordering integration — connect Tavily's supplier links to actual purchasing APIs (RockAuto, AutoZone) so mechanics can order parts directly from the dashboard Predictive maintenance — use the Neo4j knowledge graph to proactively alert customers when their vehicle profile matches patterns that historically lead to failures Multi-language voice support — VAPI supports multiple languages; MechAI should serve Spanish, French, and Arabic-speaking customers out of the box