What You'll Do
-Improve existing supervised fraud detection models (primarily XGBoost) through feature engineering, tuning, and evaluation
-Design and implement unsupervised and semi-supervised anomaly detection models (e.g., ECOD, Isolation Forest, clustering-based approaches)
-Explore ensemble strategies combining anomaly signals, supervised scores, and business rules
-Analyze fraud patterns in FNOL (First Notice of Loss) and downstream claim lifecycle data
-Develop new model concepts for emerging fraud patterns (temporal anomalies, behavior shifts, structural inconsistencies)
-Evaluate bias, stability, and confidence of fraud scores and support threshold calibration
-Partner with engineering to productionize models in AWS + Snowflake environments
-Support customer-facing analysis and explanations, including Japanese-language discussions when needed
-Document findings, assumptions, and tradeoffs clearly for technical and non-technical audiences
Required Qualifications
-5+ years of hands-on experience in data science or applied ML, preferably in fraud, risk, or anomaly detection
-Strong experience with XGBoost (or similar tree-based models)
-Practical experience with unsupervised anomaly detection techniques
-Advanced Python skills (pandas, numpy, scikit-learn, ML pipelines)
-Experience working with large tabular datasets in production environments
-Strong SQL skills; experience querying and modeling data in Snowflake
-Familiarity with AWS (S3, IAM, compute services; SageMaker a plus)
-Fluent Japanese and English (spoken and written)
-Comfortable working with partially labeled, noisy, or evolving datasets
Nice to Have
-Experience in insurance claims, auto insurance, or financial fraud
-Experience combining supervised + unsupervised signals into composite risk scores
-Knowledge of model explainability (SHAP, feature importance analysis)
-Familiarity with data pipelines (DBT, ELT workflows)
-Experience supporting regulated or compliance-sensitive environments
-Prior work with bias/fairness analysis in ML models