Flying Club Invoice Automation

The Problem

A small flying club treasurer was spending hours each month manually transcribing paper flight slips into Quicken for invoicing. Each slip contained handwritten pilot names, Hobbs meter readings, destinations, and charges — and every slip meant:

• Squinting at handwriting
• Calculating flight time (Hobbs out → Hobbs in)
• Multiplying by hourly rate
• Manually typing into accounting software

With 30-50 flights per month, this meant 5-10 hours of tedious data entry — time the volunteer treasurer didn't have.

Pain points

• Handwritten data — Varied penmanship, sometimes illegible
• Manual calculations — Easy to make arithmetic errors
• No digital trail — Paper slips get lost, damaged, or misfiled
• Context switching — Jumping between paper, calculator, and software

The Intervention

We built a mobile-first app that lets the treasurer snap a photo of a flight slip and get structured data back instantly.

How it works

1. Snap or upload — Take a photo or drag-and-drop on desktop
2. AI extraction — Vision model reads all fields (pilot, aircraft, times, route)
3. Review & edit — Verify extracted data, fix any errors
4. Export to Quicken — One-click CSV export in the exact format Quicken expects

Key insight: Train the model where to look

During testing, we discovered the AI was grabbing the wrong "name" — it found a signature field instead of the pilot name field. This highlighted the importance of:

• Field-specific prompts — Tell the model exactly what to look for and where
• Negative examples — "Don't extract from the signature line"
• Validation rules — Pilot name should match known member list

Architecture

Photo → Vision AI → Structured JSON → Review UI → Quicken CSV
          ↓              ↓
    GPT-4o/Gemini   Field validation

Stack

• App: React Native + Expo (web, iOS, Android)
• AI: GPT-4o Vision, Gemini 2.0 as fallback
• Backend: Supabase Edge Functions
• Export: Quicken-compatible CSV format

The Outcome

Before → After:

• Time per slip: 10 minutes → 10 seconds
• Monthly time: 5-10 hours → 15-30 minutes
• Error rate: ~5% (calculation mistakes) → under 1%
• Data retention: Paper in a box → Searchable digital history

Unexpected benefits

• Historical analysis — "Who flew the most this year?"
• Faster reconciliation — Match slips to payments instantly
• Member self-service — Pilots could eventually log their own flights

Key Learnings

1. Start with the export format — We built backward from Quicken's CSV spec, ensuring perfect compatibility
2. Handwriting varies wildly — Multi-model approach (try GPT-4o, fall back to Gemini) handles edge cases
3. Human review is a feature — 95% accuracy + quick human verification beats 99% accuracy that takes 10x longer
4. Small clubs have big problems — Volunteer-run organizations desperately need automation

*Ongoing cost is approximately $0.02-0.05 per slip for AI API calls.

Engagement type: Agent Prototype → Production app
Timeline: 1 week from concept to working app
Demo: flight-log-psi.vercel.app