home / case studies / Indian NBFC with digital lending products (composite)
AI / ML

MLOps Platform for Credit Scoring and Collections Models

We built a governed MLOps platform that takes credit and collections models from notebook to production with full lineage. Model refresh cycles dropped from months to weeks, with clear audit trails for regulators.

Indian NBFC with digital lending products (composite) · AI / ML

MLOps Platform for Credit Scoring and Collections Models

Challenge

Balancing data-scientist flexibility with model-risk governance, and keeping training/serving features consistent. The shared feature store and point-in-time joins removed most training/serving skew.

Approach

The platform uses MLflow for experiment tracking and model registry, Feast for a shared feature store, and Kubeflow Pipelines on Kubernetes for training and batch scoring. Real-time scoring runs behind a FastAPI/KServe service with canary releases. We added drift monitoring (population stability and feature drift), SHAP-based explanations stored with every decision, and approval gates so risk teams sign off before a model is promoted.

Outcome

Model release cycle shortened from quarterly to every few weeks Every decision traceable to model version, features and explanation Illustratively ~25–40% less analyst time spent on deployment and reruns Early drift alerts before approval-rate or delinquency surprises Reusable templates for new lending products

✓Model release cycle shortened from quarterly to every few weeks
✓Every decision traceable to model version, features and explanation
~25–40%Illustratively less analyst time spent on deployment and reruns

Problem solved

Models were trained by individual analysts, deployed by hand, and hard to reproduce. Audit queries about why a loan was rejected took days to answer.

What we built

The platform uses MLflow for experiment tracking and model registry, Feast for a shared feature store, and Kubeflow Pipelines on Kubernetes for training and batch scoring. Real-time scoring runs behind a FastAPI/KServe service with canary releases. We added drift monitoring (population stability and feature drift), SHAP-based explanations stored with every decision, and approval gates so risk teams sign off before a model is promoted.

Benefits

Model release cycle shortened from quarterly to every few weeks Every decision traceable to model version, features and explanation Illustratively ~25–40% less analyst time spent on deployment and reruns Early drift alerts before approval-rate or delinquency surprises Reusable templates for new lending products

MLflowFeastKubeflowKubernetesFastAPIKServeSHAP

Talk to us