Engineering case studies

On a platform delivering digital goods within seconds, a stolen card meant losing the product and paying a chargeback fee. Fraud could be reported up to three months after purchase. The existing hand-written rules engine failed to catch most fraudulent orders.

Score every order at checkout, before and after payment, on extremely imbalanced data whose true label can arrive up to 3 months later. Keep manual reviews within capacity without reducing conversion.

Five years of iteration in production: nearly 200 real-time features (rolling fraud rates, behavioral time windows, circular statistics); models that evolved from a random forest to an MLP, then to XGBoost with dual-model serving; automated retraining with quality gates, including a cap on manual-review volume before deployment; and model hot-swapping via pub/sub.

An online retailer of digital goods received part of its inventory as scanned sheets holding many activation keys. Staff cropped each key manually, taking several minutes per image.

Identify every key in supplier images with different rotations, noise, and surrounding text, without an existing dataset, then integrate the tool into the back office for nontechnical staff.

A semi-automatically annotated dataset of more than 20,000 keys; a custom Faster R-CNN detector later migrated to YOLOv5; post-processing calibrated to the expected number of keys; and a human-validation interface that feeds verified corrections back into the training data.

An online retailer of digital products monitored prices using supplier lists in free text. Each source mixed quantities, currencies, platforms, regions, and abbreviations in a different format.

Assign each supplier line to the correct catalog product. Two editions of the same product, such as “Edition 2021” and “Edition 2022”, differ by one character but must never be confused. An uncertain line must remain unmatched rather than receive an incorrect match.

A matching engine using domain-specific text normalization (~90 rules), structured attribute extraction, cosine-similarity vector search, and a strict consistency filter. A reference set of more than 3,000 real supplier lines guards against regressions; each catalog update triggers a hot reload across all replicas.