Problem
BlaBlaCar's Community Relations team checked every profile picture by hand against a fixed rule set. The platform received more than 19 million pictures in 2022, about 53k per day, and roughly 25% were refused. Manual work took roughly 12 full-time equivalents (FTE). Approval was also slow at peak times; normally it took a few hours.
Example from the article: during the French strikes in December 2019, the approval delay for new pictures approached 3 days.
Solution in brief
Generic steps (face detection, celebrity detection) use ready-made tools. One BlaBlaCar-specific classifier, trained on past moderation labels by transfer learning, encodes the rules. Pictures with a confident score are accepted or refused automatically. Pictures in the gray zone, where accepted and refused are about equally likely, go to humans, as does a random 5% of all pictures.
Questions this case answers
- How to automate a human review queue with a classifier while sending low-confidence cases to people and keeping an unbiased stream of labels for monitoring
- How to split a content-screening pipeline into generic steps bought off the shelf and one policy-specific model trained on historical decisions
- How to automate user-photo moderation without lowering the quality standard?
- How to decide which model predictions can be applied automatically and which must go to human moderators?
Tools and methods
- face-detection (open-source Python library) — finds faces and crops; no face or more than one face means refuse.
- AWS Rekognition — paid API that detects celebrity pictures, which are refused in countries where this is common.
- InceptionResNetV2 pre-trained on ImageNet (Keras) — frozen feature extractor, with two trainable fully connected layers on top.
- Gray-zone routing and 5% random sample — humans label uncertain cases and give unbiased labels for monitoring.
Google or a one-time email code. No password.
