ML-powered fraud detection engine scoring every incoming claim in real-time with 88% accuracy, routing high-risk cases to SIU investigators with automated evidence packs — preventing $2.8M in fraudulent claims annually.
Assured General's rules-based fraud system was flagging only 4.2% of fraudulent claims — and generating a 78% false positive rate that buried the SIU team in low-quality referrals. Estimated annual fraud leakage was $3.1M. The system was essentially a coin flip on fraud detection, consuming SIU resources while missing the majority of actual fraud.
We built FraudGuard in 12 weeks: an ML ensemble combining XGBoost and a graph neural network, serving real-time scores in under 200ms, with SHAP-generated natural language explanations for every flagged claim — transforming SIU from reactive firefighting to targeted, high-conviction case building.
The existing system caught fewer than 1 in 25 fraudulent claims. Fraud was being paid routinely — the SIU team didn't know what they weren't catching, only what occasional tips surfaced.
Fraud represents under 1% of all claims. Standard ML training on this distribution produces models that simply predict 'not fraud' for everything and achieve 99% accuracy on paper — while catching zero actual fraud.
The claims processing system couldn't be held waiting for a fraud score. Any ML solution had to return a score within 200ms at the 95th percentile — ruling out heavy ensemble models running in serial.
SIU adjusters won't act on a black-box score. They need to understand why a claim is flagged to build a case and present it to management. A model without explainability would be rejected in practice regardless of accuracy.
The existing system's false positive rate meant 78 of every 100 SIU referrals were legitimate claims wrongly flagged. SIU was spending the majority of their time clearing false alarms rather than investigating real fraud.
New claimants had no behavioural history, making ML scoring difficult. Fraudsters exploited this by creating new identities for each claim — the system needed to score new claimants from network signals, not just individual history.
An XGBoost model for individual claim features combined with a Graph Neural Network capturing claimant relationship networks — shared addresses, phone numbers, repair shops, and legal representatives. The GNN surfaces fraud rings invisible to individual-claim scoring.
Applied SMOTE oversampling to the minority fraud class plus cost-sensitive learning that weights misclassifying fraud 20x more than a false positive. Threshold tuning calibrated the precision/recall tradeoff to achieve the target 20% false positive rate.
The FastAPI ML server runs in-process with the claims system, eliminating network latency. Claimant features are pre-computed and cached in Redis on any claimant account activity — score requests hit the cache, not the database, achieving p95 under 140ms.
Every flagged claim includes a SHAP explanation translated into natural language: 'This claim is flagged because the repair shop appears in 14 prior fraud cases, the claim value is 3.2x the typical for this vehicle model, and the claimant filed 3 claims in the past 18 months.' Adjuster conviction rate on SHAP-explained referrals is 84%.
New claimants are scored on their network connections — if they share a phone, address, IP, or associated professional with known fraud entities, that signal is captured even without personal claim history. This closed the new-identity fraud vector.
An automated quarterly retraining pipeline pulls the latest 12 months of adjudicated claims, retrains the ensemble, validates against a holdout set, and promotes to production if performance exceeds the current model — keeping the system current as fraud patterns evolve.
Live in production at Week 12 with 100% of incoming claims receiving a fraud score before adjuster assignment. First 30 days: fraud detection rate at 82%, false positive rate at 21%. SIU team processed their first 3 high-confidence fraud cases in week two.
Fraud detection rate stabilised at 88.3%, false positive rate at 19%. SIU team processing 3x fewer referrals with 4x the conviction rate. $2.8M in annualised fraud prevention confirmed by actuarial analysis.
First quarterly model retraining completed. GNN analysis identified a coordinated fraud ring of 23 related claimants — a $420K exposure — surfaced by relationship graph signals invisible to individual-claim scoring. SIU successfully prosecuted 8 cases from the ring detection.
"FraudGuard went from concept to production in 12 weeks and immediately outperformed our rules engine by a factor of 20. We went from catching 1 in 25 fraud cases to catching more than 22 in 25 — while our SIU adjusters are working smarter, not harder. The SHAP explanations are a game-changer for case building and management buy-in."
Ready to modernise, scale, or build from scratch? Let's talk about your project.