Back to industries
Banking & Financial Services
Real-time transaction fraud detection
Client · European retail bank · 15M+ customers
Design and industrialization of a real-time scoring engine combining ML and business rules to detect fraud on cards and instant transfers — with a 40% reduction in false positives.
01
Challenge
- Existing fraud models relied mostly on static rules, poorly adapted to new fraud patterns (phishing, SEPA Inst. transfer fraud).
- High false-positive rate (>85%) generating customer friction and back-office cost.
- Inference latency too high to block transactions before validation (SEPA Inst. constraint: <10s end-to-end).
- Difficulty explaining model decisions to regulators and customers (ACPR, GDPR).
02
Solution
- Real-time scoring architecture on Kafka + feature store (Feast) with sub-80ms inference.
- Ensemble model (Gradient Boosting + Graph Neural Network on the beneficiary network) trained on 3 years of history.
- SHAP explainability layer providing analysts with the top features behind every alert.
- Continuous learning loop with weekly re-labeling by the anti-fraud team.
- Compliance by design: pseudonymization, audit trail, ACPR/EBA-aligned model documentation.
03
Business impact
- −40% false positives on cards, freeing 12 FTE in the anti-fraud back-office.
- +18% additional fraud cases detected vs. previous rules.
- ~€8.5M/year estimated savings on net fraud losses.
- Full industrialization in 9 months with a mixed data scientist / risk officer team.
Other industries
Insurance
Health underwriting automation with GenAI
Retail
Local assortment personalization & dynamic pricing
Energy
Wind production forecasting & market arbitrage
Manufacturing
Predictive maintenance on F&B production lines
Logistics
Route optimization & demand forecasting
Healthcare
Oncology patient pathway optimization
Let's identify where AI can create value in your business.
Book a 30-minute introductory call with a founder.