Designed a knowledge graph-based GCN model integrating CLIP text embeddings with semantic predicate-level relationships (including contrastive ones) to classify images without labeled training data; incorporated attention-based edge features and a custom prototype refinement loss and an MLP-based classifier head for final prediction.
Designed a custom LSTM-based attention model for abstractive summarization on domain-specific dialogues (SAMSum dataset). Integrated reinforcement learning techniques (SCST, PPO) to improve fluency and context retention. Developed a modular training pipeline with a custom tokenizer and lightweight RL wrapper, no transformers required. Achieved 5.42% BLEU score improvement using PPO, with <3-hour training on modest hardware.
AI-powered posture correction tool using computer vision, ML, and real-time feedback algorithms.
Developed a job market forecasting tool applying ML models for trend analysis and skill demand insights.
Developed a scalable, responsive web app with Google OAuth, state management, and payment integration.
Designed and developed a full-stack expense tracking app with secure user authentication and automated reports.
Developed a predictive language model using R Shiny and HC Corpora.
Performed EDA on storm data using statistical modeling and data visualization.
Built a Java-based ATM system with user authentication and transaction history tracking.
Research paper on using convolutional neural networks to detect pedestrian distractions for improved safety.
Published in: 2023 International Conference on Control, Communication and Computing (ICCC)