Biradar Srikanth

Fairness-Checker required a Python-native ML microservice to communicate with a Java Spring Boot backend for real-time scoring.

Initial attempts to tightly couple the ML logic inside the Spring Boot app created deployment bottlenecks — Python dependencies conflicted with the Java build pipeline.

Deployed FastAPI as an independent microservice communicating over REST. Each service became independently deployable and version-controlled, resolving the dependency conflicts.

Separation of concerns at the service level is not just a design principle — it directly affects deployment velocity and team autonomy.

The Customer Churn dataset had a significant class imbalance — far fewer churned customers than retained ones, causing the model to underpredict churn.

A baseline Random Forest model trained on raw data achieved decent overall accuracy but had a high false-negative rate — missing actual churners.

Applied SMOTE oversampling to balance the training set and tuned the classification threshold, reducing the false-negative rate by 15% while raising accuracy to 82%.

Accuracy alone is a misleading metric for imbalanced datasets. Domain-specific metrics (recall, false-negative rate) matter more than headline accuracy.

The Transaction Reconciliation Engine needed to support new discrepancy types as business rules evolved, without requiring rewrites of the core matching logic.

An initial if-else chain for different discrepancy types became unmaintainable after the third rule addition — every new rule required modifying the core validator.

Refactored to the Strategy pattern — each discrepancy type (amount mismatch, duplicate, missing entry) became a plug-in rule class. New rules could be added without touching existing code.

The Strategy pattern turns an unmaintainable conditional chain into a plug-in architecture. Design for extension from day one.

Fairness-Checker needed role-based access control for Engineer, Manager, and Admin views — each with different data visibility.

An initial session-based approach created sticky-session dependencies that complicated horizontal scaling of the Spring Boot backend.

Implemented stateless JWT authentication via Spring Security filter chain. Roles are encoded in the token, eliminating server-side session state entirely.

Stateless auth is not just a best practice — it is a prerequisite for horizontally scalable microservice architectures.