Hey There, I’m Rishikesh Sivakumar

Research Project - Dr.Anita Penkova (USC Big Data in Biotransport Center)

Paper published in "Engineering Applications of Artificial Intelligence" journal (Elsevier)

• Developed a comprehensive hybrid model that integrates Vision Transformers and CNN blocks to detect glaucoma from retinal scan images.
• The final multi-model architecture combines both images and biomarkers using fusion techniques to better detect glaucoma.
• Our Swin Transformer makes prediction with a test accuracy of 99.4% detecting 2857 out of 2874 patients’ data correctly.

Academic Project - Dr.Xuezhe Ma

Developed a novel approach to generate legally tailored patent claims by fine-tuning Google's LongT5 transformer model on a large corpus of patent data.

• Fine-tuned the LongT5 transformer on 16,000 patents using PyTorch, enabling domain-specific adaptation for legal patent claim generation.
• Integrated Low-Rank Adaptation and Parameter-Efficient Fine-Tuning techniques to reduce trainable parameters and improve computational efficiency.
• Outperformed baseline models with a 4-point and 1.5-point increase in BLEU score over fine-tuned T5 and GPT-3.5, and 10-15 point higher ROUGE scores compared to traditional approaches.

Academic Project - Dr.Dani Yogatama

Created an Emotion Detection Diary Application using Transformers and custom CNNs to analyze both texts and images providing tailored mental health recommendations based on user entries.

• Developed facial image emotion classifier powered by VGG16, trained on FER2013 dataset, achieving 90.33% accuracy rate in classifying 7 emotions.
• Fine-tuned RoBERTa on GoEmotions dataset for multi-label classification in text across 14 emotions, achieving 10% performance improvement over baseline models.

Research Project - Dr. C. Jothi Kumar

Paper presented in "Eighth International Conference on Information System Design and Intelligent Applications" and published in Springer

• Developed an automated navigation system through Computer Vision to assist the visually impaired people to navigate their environments independently.
• Used Object Detection (YOLO) and Depth Estimation to detect the objects in front of the visual aid system.
• Used a voice alert command system with PYTTSX3 to notify the user if an object is closer than 6 feet from the visually impaired individual.

HackRx Hackathon - Bajaj Finserv

Finalist (Top 30) of Pan-India Hackathon HackRx conducted by Bajaj Finserv for their problem statement SNAP Recommendation System

• Engineered a graph-based recommendation engine using Girvan-Newman Algorithm, identifying communities of frequently co-purchased items from Amazon’s SNAP dataset, improving product suggestions through hierarchical community detection on directed subgraphs.
• Enhanced e-commerce strategies by developing modular recommendation systems for Books, DVDs, Videos, and Music-CDs, leveraging graph analytics to recommend co-purchased items.
• Boosted product visibility for retailers by integrating transaction data and meta-data features, enabling prediction of top-level and second-level co-purchases, effectively prioritizing items for targeted recommendations and sales growth.

Mindspark 2.0 Hackathon - University of Southern California

Secured Third Place in Mindspark 2.0 Hackathon hosted at University of Southern California.

• Engineered a custom CNN architecture for waste overflow detection, achieving a test accuracy of 80% by leveraging data augmentation techniques for real-world generalization.
• Implemented an end-to-end workflow including data scraping, preprocessing, and augmentation using KNIME and Python, processing large datasets to optimize model performance and real-world deployment.
• Developed an interactive GUI for end-user engagement, allowing real-time waste bin classification as clean or overflowing, with potential applications for city-wide waste management optimization.