Selected Projects
Video Question Answering Capstone Project
PES University | May 2022 - Dec 2022
- Devised a multi-modal video question-answering model on the TVQA dataset within limited resources.
- Optimized the model to use just 17.5M trainable parameters, achieving 68.07% test accuracy; comparable to state-of-the-art models of that period with 100M parameters.
- Demonstrated efficiency in a multi-modal context.
Anomaly Detection in Credit Card Transactions
Centre for DS and Applied ML lab, PESU | Aug 2020 - Feb 2022
- Addressed fraudulent transaction detection on a highly imbalanced Credit Card dataset (0.172% fraud rate).
- Designed two novel solutions: the BEUD hybrid model (Autoencoder + Siamese) improving over Autoencoder, and a Custom Binary Cross Entropy (CBCE) loss function outperforming standard Binary Cross Entropy.
- Attained a 10% test recall improvement with BEUD and 2% improvement with CBCE, mitigating false negatives and addressing limitations of baseline approaches.
Celebrity Recognition Group Project
PES University | Jan 2022 - May 2022
- Implemented celebrity recognition in video footage, addressing redundant frames and low-quality facial images.
- Designed a multi-stage pipeline: key frame extraction, face extraction, super-resolution, and encoding-based recognition (using Siamese Network).