Travel · Enterprise AI · Regional OTA
Document intelligence that cleared a 40,000-file backlog
A multimodal model reads visas, passports, and booking forms, extracts the fields, and routes only the uncertain ones to a human.
PythonVision modelAirflowPostgres
-70%
processing time
Timeline
3 months
Team
3 engineers + 2 ML engineers
The challenge
What was breaking
- 01A 40,000-file backlog of manual document review.
- 02Slow, error-prone data entry during peak season.
- 03No way to scale review without scaling headcount.
The approach
How we tackled it
- 01Trained multimodal extraction across document types and languages.
- 02Added a confidence layer that routes low-certainty fields to humans.
- 03Built straight-through processing for everything above threshold.
The solution
What we shipped
- 01Auto-extraction of every required field.
- 02Human review reserved for the genuinely uncertain.
- 03Straight-through processing for the clear majority.
In pictures
A closer look
The results
40k
backlog cleared in 3 weeks
-70%
processing time
99.2%
field accuracy
80%
straight-through
"We cleared a backlog we'd stopped believing we'd ever clear. The team now reviews exceptions, not everything."
Related work