Built an LLM-powered assessment and feedback system using GPT-4 and retrieval-augmented generation (RAG) to automate rubric-aligned report generation, processing 10,000+ submissions with 95% acceptance rate.

Designed scalable, production-grade GCP data pipelines leveraging BigQuery, App Engine, and Cloud Scheduler to ingest, clean, and transform 10K+ daily video and form records from VideoAsk and HubSpot for downstream analytics and reporting.

Refactored monolithic services into modular Java/ Python microservices with event-driven architecture, enabling real-time filtering and reducing latency by 70%.

Trained and evaluated ML models (BERT, XGBoost, Random Forest, LSTM) to flag low-effort/ off-topic responses and forecast competency growth, achieving 87% accuracy in rubric-level proficiency change prediction.

Streamlined internal reporting through CI/ CD automation with GitHub Actions and Docker containerization, cutting manual workload by 85% and adding automated quality checks to monitor output drift.

Developed secure RBAC-compliant REST APIs integrated with PostgreSQL, enforcing HIPAA-aligned access policies.

Improved API performance by 40% through profiling, applying concurrency enhancements, and optimizing database indexing for datasets.

Conducted rigorous testing, deployment, and rollout of backend services used daily by 1,000+ medical staff across clinics.

Developed a logistic regression propensity scoring pipeline using clickstream, purchase-cycle, and geo-demographic features to identify high-intent merchants and improve campaign ROI.

Conducted A/B testing and funnel analysis across email, SMS, and ad channels using Python and statsmodels, improving conversion rates by 25% and guiding budget allocation.

Refactored and optimized 1,000+ lines of SQL (CTEs, window functions, nested queries), reducing report runtimes by 60% and improving dashboard freshness.

Trained and evaluated ML classifiers (SVM, XGBoost, Random Forest, ANN) using 5-fold cross-validation, ROC-AUC curves, and confusion matrices to compare precision/recall trade-offs.

Selected an ANN (TensorFlow/Keras) after comparative evaluation showed stronger generalization on non-linear relationships, achieving 92% prediction accuracy with improved minority-class recall.

Tuned the neural network (hidden layers, activations, dropout) using randomized search with early stopping and learning-rate decay, reducing overfitting and cutting training time by ~30%.