|

Designed and implemented a real-time network anomaly detection system by integrating Snort IDS with multiple unsupervised ML models including Isolation Forest, LOF, One-Class SVM, K-Means, GMM, and Elliptic Envelope.

Conducted a comparative performance analysis across six anomaly detection techniques, demonstrating superior detection accuracy using ensemble decision-making, with Isolation Forest and LOF achieving the highest results.

Built a low-latency Streamlit-based dashboard to process live network traffic from Snort logs, achieving high accuracy and recall suitable for real-time intrusion detection deployments.

Developed and evaluated reinforcement learning–based autonomous control policies for spacecraft landing under uncertain dynamics using OpenAI Gymnasium’s LunarLander-v2 environment.

Implemented a unified PyTorch framework to compare DQN, Double DQN, Dueling DQN, and PPO across multi-seed experiments, analyzing convergence speed, reward stability, and computational efficiency.

Demonstrated that PPO converges 4× faster than value-based methods while Dueling DQN achieves the highest peak reward, highlighting trade-offs for real-world high-risk autonomous control systems.

Designed a 100 W wireless power transfer system using inductive and resonant power transfer with a Series–Series compensated topology to achieve high efficiency and safe power delivery.

Implemented a complete transmitter–receiver architecture, including SMPS, DC–DC conversion, rectification, filtering, and a Battery Management System (BMS) for compliant and reliable battery charging.

Optimized State of Charge (SOC) control and battery lifecycle management, validating system performance through simulation and prototype testing to ensure durability and efficiency.

Hybrid unsupervised intrusion detection combining deep autoencoders with adaptive z-score filtering for dynamic anomaly thresholding.

Achieved 91.4% accuracy and improved performance by 8–15% compared to Isolation Forest, One-Class SVM, and classical statistical approaches.

Evaluated on 125,973 diverse network traffic samples including DoS, Probe, R2L, and U2R attacks.

Real-world deployment across enterprise, academic, and cloud networks showing 67% fewer false positives and SIEM integration with Splunk, QRadar, and ArcSight.

Built a collaborative search engine that re-ranks results based on user behavior including scroll depth, time spent, and click-through patterns.

Implemented extractive summarization using MIS and eigenvector centrality to generate concise document summaries.

Enabled knowledge sharing so users in a group can view and benefit from each other's interactions and ranking signals.

Demonstrated improvements in search relevance and faster retrieval for multi-user research workflows.

Developed a deep-learning-based image-to-audio conversion system providing real-time audio descriptions for visually impaired users.

Used CNN-based feature extraction combined with speech generation for object recognition and scene understanding.

Built a mobile-friendly interface enabling users to easily capture images and receive audio interpretation.

Validated through user studies showing high accuracy and major improvements in usability and daily navigation.