Aircraft Journey Summary Extractor

01 /Overview

Journey summary forms are partly handwritten and full of aviation abbreviations, which traditional OCR reads poorly.

02 /What I built

• Gemini 2.0 Flash vision chosen over OCR plus NLP, AWS Textract and GPT-4 Vision, with the trade-offs documented
• Prompt with aviation expertise and abbreviations (FWD, GND, A/C, DD, MEL)
• JSON validation: required fields, integer types, clean-up of model output
• Graceful batch processing: one bad form does not block the rest
• Estimated cost of about $0.001 per form

03 /Outcome

A single-step extractor that returns validated JSON (aircraft, registration, route, crew, fuel, load, defect message) in about two to three seconds per form.

04 /What I learned

Domain context in the prompt was the biggest lever: telling the model it is a maintenance technician and listing abbreviations lifted handwriting accuracy from about 60% to about 95%.