Cash-Audit Fraud Detection
Detects shortfalls and unusual behavior in the cash-ups of hundreds of stores, combining three machine-learning approaches.
The problem
Manually auditing cash counts across hundreds of stores doesn't scale: anomalies surface late or get lost among thousands of transactions.
What I did
Built the whole system: supervised models (Random Forest and XGBoost with hyperparameter search, feature selection and cross-validation), K-Means clustering with cosine similarity and Isolation Forest for anomalies, plus a conversational layer with LangChain. Versioned trained models with performance metadata — pragmatic MLOps — and wrote the deployment guide.
The outcome
In production in two countries, integrated into the internal AI portal.
How it detects
- Cash-ups from hundreds of storesInput
- Supervised modelsRandom Forest & XGBoost
- K-Means clusteringCosine similarity
- Isolation ForestUnlabeled anomalies
- Conversational layerLangChain
- Detected anomaliesOutput
Three approaches over the same cash audits: the supervised models catch what is already known, clustering groups behaviors, and the anomaly detector flags what resembles nothing. The conversational layer lets you ask why.
Stack
- Python
- scikit-learn
- XGBoost
- Streamlit
- LangChain
- Plotly
- Docker
Blue Star Group internal project — the code is not public.
Next project: TeamSync →