Developed a loan-default prediction model for Home Credit applicants, training a LightGBM model on 307,511 entries and achieving a ROC-AUC of 0.68. Fixed a calibration mismatch by deriving thresholds from the training-set score distribution. Built a FastAPI-based Borrower API with email-OTP auth and SHAP attribution, and a Lender Console for review queues and PSI drift monitoring.

Constructed a streaming anomaly-detection service to monitor 27-feature traffic windows, training on chaos-injected telemetry and achieving an F1 score of 0.914, with PR-AUC 0.980 and ROC-AUC 0.987. Integrated SHAP TreeExplainer into the Kafka consumer for per-feature attribution in alerts, and deployed as 13 Docker containers with Prometheus/Grafana observability for model-health and PSI drift monitoring.

Executed a comprehensive benchmark comparing plain LLM, vector RAG, and GraphRAG over a 432-article TigerGraph knowledge graph, improving LLM-as-Judge accuracy from 64% to 71%. Identified and corrected a 13x token cost under-report by scraping Docker logs, and raised accuracy by 10 points through disabling an internal re-ranker. Implemented 3-vote majority consensus and an adaptive 2-hop graph-traversal fallback to achieve 92.9% judge accuracy and 0.89 BERTScore F1.

For JC Decor, I enhanced their custom ERP Lite system by resolving critical filtering issues for designer and machine fields. This streamlined operations, improved system logic, and laid groundwork for future integrations.

For YOS SPORTS HEALTH SPECIALISTS' YOS 2.0, I optimized their Analytics & Dashboards platform, significantly improving package and calendar slot API performance. This enhanced data retrieval and added support for diverse correlation type representations.

An outdated UX/UI created a non-seamless experience for YOS Sports Health Specialists' customers and staff. We delivered YOS 2.0, a complete platform revision including a redesigned UK booking flow and new management systems, resulting in an intuitive, engaging journey and streamlined internal operations.

Indian gas trading suffered from delays, errors, and zero customer visibility due to manual calls and WhatsApp coordination. Silendr digitized the entire lifecycle with a mobile-first ordering platform, integrating with ERP/CRM, delivering a seamless, transparent, and personalized customer experience.