AI face-recognition frauddetection
Face-recognition model implemented for fraud detection, with an evaluation harness measuring accuracy before rollout.
- Client
- Confidential
- Type
- AI
- Year
- 2025
- Delivered
- AI engineering · Model evaluation · Pipeline build
- Location
- UAE
- Method
- Evaluation-first
- Type
- AI pipeline
- Rollout
- Shadow mode first
01 /
The challenge
Manual checks were missing fraudulent entries that looked identical on paper, and there was no way to measure how often the system was wrong. The client also needed a clear answer to one question: when the model is not confident, what happens next?
02 /
What we did
- 01
Built the evaluation harness first — labelled data, measured accuracy, a clear confidence threshold — before any model touched production, so the accuracy claim is tested rather than asserted.
- 02
Shipped a pipeline with a human-escalation path for low-confidence matches, so edge cases are reviewed by a person rather than silently auto-rejected or auto-approved.
- 03
Documented the accuracy envelope clearly: what the model catches, what it misses, and where it must not be trusted — so the client knows exactly what they are deploying.
- 04
Designed the integration so it slots into the existing check workflow rather than replacing it, with a dashboard showing match rates, escalations and overrides.
03 /
Outcome
A measured system, not a demo. Accuracy is reported against a harness the client can rerun, edge cases go to a human, and the integration works inside the existing check process rather than beside it. The client can answer, with numbers, how well it is working.
Built with
- Python
- TensorFlow
- OpenCV
- FastAPI