Fintech Transforms Risk Decisioning
with ML and Azure MLOps
Customer Background
A digital lending platform active in Southeast Asia and Africa, servicing over 3 million borrowers through BNPL and micro-loans. With a lean operations team and aggressive scale goals, the client’s business depended on accurate risk models and fast credit approvals. Their existing credit system relied on basic rule sets and slow manual reviews, resulting in default spikes and approval delays.
Challenges
The scoring system failed to capture user behavior signals, resulting in 18% default rate spikes in new user groups. Loan approvals took hours due to fragmented data and manual overrides. The client lacked a way to experiment, version, or explain ML models. Regulatory bodies demanded transparent, explainable decisioning and audit logs. High latency and model drift plagued production deployments, leading to unstable user experiences and rejections.
Solutions
We designed and deployed a modular credit scoring pipeline using XGBoost, trained on 500K+ historical repayment records. Data was ingested and transformed via Azure Data Factory, with model lifecycle managed in Azure ML using versioned endpoints. SHAP values powered explainability and compliance documentation. The scoring model was deployed as a real-time API through Azure Functions with Cosmos DB-backed logs. Drift monitoring and auto-retraining workflows were set using MLflow and Azure Pipelines. Performance monitoring dashboards were built in Grafana.
LLM Integration: Used OpenAI’s GPT-3.5-turbo fine-tuned on product metadata, customer reviews, and query history
Vector Search Engine: Implemented semantic search with FAISS, training custom product embeddings using SentenceTransformers
Personalization: Leveraged real-time behavioral signals from Databricks Delta Live Tables for context-aware results
Deployment & APIs: Served the assistant via scalable REST APIs on Azure Kubernetes Service (AKS) with Redis caching
Fallback Mechanism: Introduced confidence-based fallback to legacy keyword search when GenAI confidence was low
Analytics Layer: Power BI dashboards monitored usage, conversion lift, query quality, and revenue attribution
This architecture allowed fast, scalable GenAI experiences with performance tuned specifically for mobile environments and regional traffic surges.
Benefits
- 38% improvement in product discoverability
- 27% increase in conversion rate on mobile
- 45% lower query latency using edge caching
- 60% of users preferred GenAI assistant over keyword search
- 22% reduction in cart abandonment rate
- AI-driven merchandising insights improved catalog performance