Developed computer vision pipelines for sports analytics on broadcast video, enabling player tracking and event detection to extract performance metrics. Deployed a low-latency conversational search system over 500+ technical components using LangChain and FastAPI, achieving <150ms response times and 91% answer accuracy via LLM-as-judge evaluation.

.Built end-to-end health records policy and fraud detection and risk-scoring ML systems using XGBoost, LightGBM, and time-series anomaly detection, reducing false positives by ~25%. Designed scalable batch and near–real-time inference pipelines processing millions of transactions daily, added SHAP-based explainability for compliance, and automated training and deployment with MLflow and CI/CD—driving a ~35% reduction in fraud losses during pilot rollout.

• Developed autonomous underwater robots with high-resolution optical sensors, reducing inspection costs by 93% and enhancing defect detection reliability through improved data collection and real-time analysis capabilities.

• Enhanced YOLOv5 model performance by applying Gaussian filtering and CLAHE, achieving 87% higher defect detection accuracy and optimizing CSPDarknet53 with FPN and PANet for 90.7% precision, 88.3% recall, and 12ms inference time.

• Delivered engaging lessons on advanced ML concepts, including Random Forests, Transformers, and Gradient Boosted Machines, enabling students to build projects like predictive models and text summarization systems.

• Mentored students on cutting-edge AI topics such as Generative AI and Recommendation Systems, incorporating real-world techniques like GPT, Collaborative Filtering, and Matrix Factorization to foster hands-on problem-solving skills.

• Conducted workshops for over 50 students, guiding them in implementing ML models such as Random Forests, ARIMA, and Transformers, resulting in a 30% improvement in their practical understanding and project outcomes.

• Assisted in research projects by developing DL architectures like LSTMs and utilizing tools like SHAP and LIME, leading to a 20% increase in model interpretability and contributing to successful completion of two key forecasting studies.