AI Credit Scoring for Thin-File Customers

The fintech's traditional scoring model relied primarily on information from credit bureaus, loan history, payment behavior, and financial product tenure. While this approach worked well for customers with established credit histories, it presented significant limitations when evaluating thin-file profiles.

As a result, some potentially creditworthy customers could receive an insufficient risk assessment simply because the traditional model did not have enough data to classify them accurately.

The risk team also needed greater transparency into model-generated predictions. Obtaining a score alone was not enough; analysts needed to understand which variables influenced the prediction and why a customer was classified within a particular risk level.

The challenge was to develop a more predictive and flexible scoring capability that could incorporate alternative data, provide explanations for its predictions, and integrate directly with the credit origination platform.

Rootstack designed and implemented a Machine Learning-based AI Credit Scoring platform capable of complementing traditional credit information with alternative data and converting it into useful risk signals for the origination process.

The solution was built as an independent service, allowing models to evolve without requiring direct modifications to the origination platform. The architecture followed a complete pipeline consisting of data sources, processing, Feature Engineering, Feature Store, Machine Learning model, XAI, and a Scoring API connected to the origination platform.

The platform integrated information from 10+ data sources, including core banking, CRM, credit bureaus, transactional systems, payment platforms, and digital channels. The data went through normalization and validation processes to identify missing values, duplicates, inconsistencies, and unexpected changes in data distribution.

From these sources, the platform generated 100+ features capable of capturing financial behavior patterns beyond traditional variables. These included indicators such as income stability, transaction frequency, balance trends, payment regularity, credit utilization, cash-flow variability, and financial product usage.

The generated variables were managed through a Feature Store to maintain consistency between training and production data. This enabled feature versioning and reuse while reducing discrepancies between model development and production environments.

Machine Learning models were implemented on top of this layer through a pipeline covering dataset preparation, feature selection, training, validation, calibration, and model versioning. Models were evaluated using credit risk metrics such as ROC-AUC, Gini, KS statistic, precision, recall, and probability calibration. The solution achieved an estimated 15–20% improvement in predictive power compared with the baseline, with performance also evaluated specifically across thin-file customer segments.

A key component was the incorporation of Explainable AI through SHAP. The platform could accompany each prediction with the main factors contributing to the result, allowing risk analysts to understand the influence of positive and negative variables on the score. 100% of predictions included XAI explanations.

Once generated, the score was exposed through a Scoring API that could be consumed directly by the origination platform. The integrated workflow covered credit application, customer validation, data enrichment, AI Scoring, risk assessment, rule evaluation, and the final outcome, which could lead to automatic approval, manual review, or rejection.

The architecture maintained a clear separation between the predictive model and deterministic business rules. The Machine Learning model could calculate a probability of default or risk category, while the decision engine remained responsible for applying formal product policies, exposure limits, eligibility criteria, and regulatory requirements.

The platform also incorporated MLOps capabilities to version datasets, features, models, parameters, and metrics. Once deployed, models were continuously monitored for data drift, feature drift, changes in score distributions, and predictive performance degradation, allowing models to be reviewed or retrained when necessary.

From an infrastructure perspective, scoring was implemented as an independent microservice, enabling horizontal scalability and reuse across web applications, mobile applications, origination platforms, and partner APIs. The integration layer incorporated authentication, authorization, auditing, and environment isolation, allowing each scoring request to be traced to the system that initiated it, the model version used, and the result generated.

The project was completed in 6 months. The solution enabled 60%+ of thin-file customers to be evaluated automatically, reduced scoring time by 80%, and achieved 100% integration of scoring into the origination workflow. Additionally, 100% of models were versioned and monitored, while the Scoring API maintained 99.9% availability.

The implementation used Python, Machine Learning, Pandas, NumPy, Feature Store, SHAP, MLflow/Model Registry, FastAPI, PostgreSQL, Docker, Kubernetes, REST APIs, Kafka, MLOps practices, and OpenTelemetry for data processing, model development and management, enterprise integration, scalability, and observability.