CATAI

Inspiration

Heavy equipment inspections are high stakes: a missed bolt, cracked hose, or “off” engine sound can lead to downtime, expensive repairs, or a safety incident. We wanted to build a fast, structured AI copilot that turns messy field observations into clear PASS / MONITOR / FAIL decisions, quantified risk, and a finalized report — and gets smarter over time.

What it does

CATrack (CATAI) is an iOS inspection assistant (single Caterpillar model MVP) that lets an inspector:

• Chat in plain English (and voice) about what they see/hear
  
• Automatically update the inspection checklist with what changed + why
  
• Generate a final inspection report (executive summary, critical findings, recommendations, operational readiness, risk score)
  
• Save the report to Supabase (archive / source of truth)
  
• Use Supermemory to remember past inspection summaries per machine, so the AI can flag recurring issues and adjust recommendations
  

How we built it

• iOS (SwiftUI): chat UI, machine picker, editable inspection sheet, archive view, capture + upload flows
  
• FastAPI backend: endpoints for starting inspections, analyzing text/voice, generating reports, sound anomaly checks, plus debug tools
  
• OpenAI: reasoning + structured JSON outputs for checklist updates and report generation
  
• Supabase: persistent storage for inspections + generated reports (and media paths)
  
• Supermemory: stores/retrieves machine-tagged inspection summaries to add “history-aware” intelligence to every new analysis
  

Challenges we ran into

• Deploying backend for real iPhone testing: localhost worked on simulator, but phones needed a public endpoint → fixed by deploying and binding correctly to platform host/port
  
• Supermemory retrieval returning empty results: learned container tag matching + response formats; built debug endpoints to verify add/search behavior
  
• Keeping UI + sheet state consistent: ensuring AI updates don’t overwrite user edits, syncing notes reliably, and making explanations visible in-chat
  

Accomplishments that we're proud of

• End-to-end loop working: start inspection → AI updates checklist → generate report → archive + memory
  
• The assistant can say “this is recurring” because it actually retrieves prior inspection history (verified in outputs)
  
• Reports are persisted in Supabase, while memory stays fast and relevant via Supermemory
  
• Built a clean, demo-ready UX: machine list, chat, sheet editing, finalize/report, archive
  

What we learned

• The best architecture is two-layer memory:
  
  • Supabase for canonical records + audit trail
    
  • Supermemory for semantic recall that improves AI decisions
    
• Deployment details matter as much as code in a hackathon
  
• AI output must be actionable: inspectors need structured updates + reasons, not just a paragraph
  

What's next for CATAI

• Expand from one-model MVP to multiple models + templates with a parts library
  
• Improve quantification (probability/time-to-failure/cost/downtime) using more data + calibrated heuristics
  
• Add stronger image-first defect detection + localization (bolts/hoses/leaks)
  
• Create a maintenance ticket workflow (export PDF / CMMS integration / shareable report links)
  

Built With

• fastapi
• openai
• postgresql
• python
• railway
• supabase
• supermemory
• swift
• swiftui
• uvicorn