Investigated PPO/DPO/GRPO for alignment in retrieval-augmented generation; benchmarked performance using Ragas metrics for faithfulness and relevancy.

Built a pipeline using Random Forest and XGBoost for financial fraud prediction.

Modeled using Ornstein-Uhlenbeck mean-reversion pro cesses and Monte Carlo simulations.

Applied Isotonic Regression and Platt Scaling for risk bucket calibration; evaluated via ROC-AUC and Kolmogorov-Smirnov statistics.

Developed taxonomy of 50+ LLMs evaluation metrics; provided task-specific recommendations and decision framework.

Built a CNN-based MER pipeline with ECG/GSR and BRR/GPR models for valence–arousal prediction (MAE ≈ 0.04, RMSE ≈ 0.057); achieved personalized emotion-aware robotic adaptation via Q-learning (r = 0.7185).

Used and fine-tune YOLOv5 and MobileNetV2 pretrained models for pedestrian detection, achieving nearly 85% model accuracy on CityPerson dataset.

Used Apache Kafka and Spark with KNN and K-means algorithms for real-time cryptocurrency behavior analysis, applying Silhouette and Elbow methods to evaluate clustering performance.

Built a hybrid LSTM + MHSA + ANN model on the OULAD dataset with significant accuracy by 80%.