Engineering intelligent systems from data.

Real-time PPE compliance detector with YOLOv8/YOLO26 and OpenCV, flagging missing helmets, vests, gloves, goggles, and boots from webcam or video feed. Detects missing safety gear across a 10-class scheme, logs violations to CSV with cooldown deduplication, and ships with a live FastAPI dashboard for webcam, video file, or RTSP stream sources.

Fine-tuned YOLOv11n for shelf analytics (health score, fill rate, restock alerts) on 44 retail shelf images with advanced augmentation at 640×640. Built a COCO-to-YOLO converter and zone-level breakdown engine. mAP50 46.4%, precision 45.2%. Deployed via REST API with ONNX export for edge inference.

Three class sentiment model using BiLSTM with Bahdanau attention on 60K Amazon reviews. Inverse frequency weighting lifted neutral class F1 from 0.00 → 0.46 and macro F1 from 0.57 → 0.63.

FraudNet - a PyTorch MLP trained on 284K transactions with SMOTE oversampling to tackle severe 1:578 class imbalance. AUC-ROC 0.964, 86.5% recall. Optimized with cosine annealing and early stopping. Deployed FastAPI REST API with real-time LOW / MEDIUM / HIGH / CRITICAL risk scoring and interactive web UI.

Trained an XGBoost churn model on 5K customers (31% churn rate) with SHAP explainability served via FastAPI. Engineered 33 features from 21 raw columns; tuned via 5-fold RandomizedSearchCV over 50 iterations. Features an executive dashboard with KPIs, segment churn rates, risk distribution, real time churn predictor with retention recommendations, and a customer risk explorer with scatter plots and paginated filtering.

PyTorch pipeline for multi label toxicity detection across six categories on the Jigsaw dataset. Systematic feature extraction and hyperparameter optimization - achieving 92% F1 score across 6 toxicity categories.

AI driven search agent with multi step reasoning and contextual memory across multi turn queries. Improved tool selection strategy reduced response latency by 40%.