FIXITAI

FixIt AI — HenHacks 2026 Submission Inspiration We walked past a leaking pipe in Gore Hall and realized — the only way to report it is to email fixit@udel.edu, type a vague description, and hope someone reads it. No photo. No priority. No feedback loop. Meanwhile, UDel's Facilities team manually reads every email, figures out whether it's plumbing or electrical, decides how urgent it is, creates a work order in IBM Maximo, and routes it to the right person. That's five human bottleneck steps before anyone even touches the pipe. Then we asked a bigger question: what if reporting didn't even require hands? What if a maintenance worker could walk through a building wearing Meta Ray-Ban smart glasses, spot an issue, and have it automatically documented — the AI sees what they see, classifies it, and dispatches the work order by the time they've taken three more steps? That vision of hands-free, AI-powered infrastructure management is what drove FixIt AI. What it does FixIt AI is a full-stack campus maintenance automation platform with three portals — students, technicians, and facility managers — powered by AI vision, voice, and real-time intelligence. For Students — Report in Seconds, Not Minutes:

Snap a photo of any campus issue — GPT-4o Vision instantly identifies the problem type (plumbing, electrical, HVAC, structural, safety hazard), assigns priority, and estimates repair cost and time Voice reporting via ElevenLabs and Web Speech API — describe the issue by talking instead of typing, the AI transcribes and classifies it alongside the photo Emergency button for critical safety hazards that triggers immediate escalation QR code scanning — scan a code posted in any building to auto-fill location data, no building dropdown needed Real-time status tracking on every report you've submitted

For Facility Managers — Intelligence, Not Spreadsheets:

Live campus safety dashboard with real-time report feeds powered by Supabase real-time subscriptions Interactive campus map showing every report as a color-coded marker — red for critical, amber for high, blue for medium, green for resolved Smart deduplication — when 10 students report the same broken elevator, the system merges them into one work order and auto-escalates the urgency score Pattern detection — "Gore Hall has had 6 plumbing issues in 90 days. Recommend preventive pipe inspection. Prevention cost: $2,000. Estimated cost if main bursts: $50,000+" AI-powered technician assignment — automatically matches the right specialist based on trade type, availability, and proximity Escalation engine — reports that aren't addressed within SLA thresholds auto-escalate to department heads Batch job processing for handling bulk historical reports QR code generator — print and post QR codes in every building so students can scan to report

For Technicians — Clear Jobs, Fast Resolution:

Mobile-optimized job queue with priority-sorted cards Full job details with AI analysis, photos, building floor plans, and suggested repair actions One-tap status updates — accept, start, complete — with photo documentation of completed repairs Push notifications for new assignments via real-time hooks

The AI Brain:

GPT-4o Vision analyzes maintenance photos and returns structured JSON: trade classification, priority level, safety concern flag, estimated repair cost, estimated repair time, confidence score, and suggested action Email ingestion pipeline — can process reports that still come in via the traditional fixit@udel.edu email, auto-classifying them the same way Automated dispatch via Gmail SMTP — the correct department gets an email with the photo, AI analysis, location, and priority before a human even looks at the queue Preventive maintenance engine that analyzes historical patterns across buildings and recommends proactive interventions

How we built it Stack: Next.js 16 (App Router) with TypeScript, Tailwind CSS, and shadcn/ui for the frontend. Supabase PostgreSQL for the database with real-time subscriptions. OpenAI GPT-4o Vision API for image analysis and classification. ElevenLabs and the Web Speech API for voice input and transcription. Nodemailer with Gmail SMTP for automated email dispatch. React-Leaflet for the interactive campus map. Deployed to Vercel with auto-deploy on every git push. Collaboration: We used Claude Code with the Remote Control feature — allowing one teammate to control the development session from their phone while the code ran on the laptop. One of us built the entire backend: API routes for analysis, reporting, email ingestion, escalation, preventive maintenance, batch processing, technician assignment, and QR code generation. The other managed the frontend portals, design system, PWA configuration, and deployment pipeline. GitHub branch strategy (feature/frontend and feature/backend merging to main) kept us in sync. Design philosophy: After the first build, the UI looked unmistakably AI-generated — purple gradients, oversized rounded corners, emoji everywhere. We deliberately stripped every generic element and rebuilt the design system from scratch following Stripe and Linear's aesthetic: flat surfaces, minimal borders, tight spacing, one accent color (UDel blue #00539F), no shadows deeper than 1px, no border radius above 6px. The result looks like a product, not a hackathon project. Meta Ray-Ban Integration (Research + Architecture): We researched the Meta Wearables Device Access Toolkit and designed the system architecture for hands-free reporting. The camera sensor on Meta Ray-Bans pipes image data to a companion mobile app via Bluetooth, which hits our /api/analyze endpoint — the same pipeline that powers the phone-based reporting. We also explored the open-source meta-glasses-api browser extension for routing voice commands through Messenger to our AI backend. The software is glasses-ready today — plug in the hardware and it works. Challenges we ran into Hotel WiFi at the venue blocked device-to-device traffic, so we couldn't share localhost between our laptops. We pivoted to Vercel for instant deployment and used GitHub as our sync layer. Claude Code's Remote Control feature had a Node.js v22 bug where the --sdk-url flag was being passed to the Node binary instead of the application. We spent 30 minutes debugging before discovering it only affected npm installations — installing the native binary alongside the npm version fixed it. GitHub's secret scanning blocked our first push because API keys were committed in our CLAUDE.md project context file, forcing us to restructure our entire environment variable handling. Getting GPT-4o Vision to return consistently structured JSON required iterative prompt engineering — early versions hallucinated repair costs or returned malformed responses. Building three separate user portals (student, manager, technician) with proper authentication flows in 6 hours meant every decision had to be fast and final — no time for deliberation. Accomplishments that we're proud of In under 6 hours, with two people, we built a production-grade platform with: three role-based portals with authentication, AI-powered image analysis that correctly classifies trade types and assigns realistic priorities and cost estimates, automated email dispatch to the right department, voice-powered reporting, QR code location scanning, an emergency alert system, a real-time safety dashboard with live campus mapping, smart deduplication that merges duplicate reports and escalates urgency, a preventive maintenance engine that detects recurring patterns across buildings, an escalation system with SLA-based auto-escalation, technician job management with assignment and completion workflows, floor plan integration, push notifications via real-time Supabase subscriptions, batch job processing for historical data, and an email ingestion pipeline for legacy fixit@udel.edu reports. The system handles the full lifecycle — from a student spotting a problem to a technician completing the repair — with AI automation at every step. We eliminated the entire human dispatcher bottleneck. What we learned Infrastructure setup eats hackathon time alive. We spent the first two hours on SSH keys, Git configuration, WiFi workarounds, Vercel deployment, Supabase provisioning, and environment variable management before writing a single line of application code. Lesson: front-load every infrastructure decision and use the simplest tools that work. AI-generated UIs have a recognizable aesthetic that experienced judges spot immediately. Deliberately designing against those patterns — stripping gradients, reducing border radius, eliminating emoji, using one accent color — transforms a project from "clearly AI-built" to "clearly product-minded." The design conversation matters as much as the technical architecture. Claude Code with Remote Control changed our workflow fundamentally. Being able to send development instructions from a phone while the code runs on a laptop meant we never stopped building — even during bathroom breaks and food runs. The future of development is untethered. Finally, we learned that the best hackathon pitch isn't "look at all these features" — it's "here's a real problem you've personally experienced, and watch us solve it in 10 seconds live on stage." What's next for FixIt AI

UDel Pilot — Partner with Facilities & Auxiliary Services to run a live pilot, feeding reports directly into their IBM Maximo system via API integration. We've already documented how their current workflow operates and where FixIt AI plugs in. Meta Ray-Ban Hands-Free Mode — Our API is already glasses-ready. Using Meta's Wearables Device Access Toolkit, a maintenance worker walks through a building wearing Ray-Bans, says "Hey Meta, report this," and the glasses capture the image, our AI classifies it, and the work order is dispatched — all hands-free. We've also explored OpenGlass ($25 DIY smart glasses) as a budget alternative for student workers. ElevenLabs Voice Assistant — Expand voice capabilities from input-only to full conversational interaction. A technician asks "What's the status on Gore Hall plumbing?" and gets a spoken response with their job queue and priorities. Multilingual support for international facility staff. Predictive Maintenance AI — With enough historical data, move from reactive to predictive. "Building X's HVAC system will likely fail within 30 days based on report frequency, seasonal patterns, and equipment age." This is the real money — preventing a $50K emergency with a $2K scheduled repair. Multi-Tenant SaaS Platform — FixIt AI was built for UDel, but the problem exists wherever buildings exist. Universities (1,000+ in the US), property management companies, corporate campuses, hospitals, municipalities. The CMMS market is worth $1.5B+ and current solutions like IBM Maximo cost $100K+/year. FixIt AI deploys in minutes at a fraction of the cost. Native Mobile Apps — Convert the PWA to native iOS and Android with push notifications, offline mode, and background location for proximity-based alerts ("You're near a reported hazard in Smith Hall").

Built With

• next