Projects

Independent Project

• Proposed a novel POSIX Loss penalizing inconsistencies in responses to semantically equivalent prompts, significantly enhancing robustness in Vision-Language Models.
• Implemented comprehensive benchmarks for medical and general-purpose VLMs, evaluating robustness across spelling errors, template variations, and paraphrasing.
• Achieved notable improvements on medical imaging datasets, increasing accuracy (up to 72.9%) and reducing hallucinations via optimized training and fine-tuning methodologies.

Independent Project

• Developed an AI-powered legal simulation system using CrewAI to model courtroom proceedings with Groq LLM agents representing judges, juries, prosecutors, and defense attorneys for historical serial killer cases.
• Implemented comprehensive output generation and storage producing detailed reports for each participant's perspective, with standardized file naming and organization for further analysis.
• Created data visualization components to analyze sentencing patterns, victim counts, and relationships between case factors using Matplotlib and Seaborn.

Independent Project

• Developed a retrieval-augmented generation (RAG) system utilizing ChromaDB for embedding storage and Groq's LLaMa 3.1 model for generating context-aware responses.
• Implemented efficient semantic retrieval leveraging embeddings generated by the all-MiniLM-L6-v2 model for improved accuracy.
• Engineered flexible document chunking strategies, balancing context preservation and retrieval precision to optimize response accuracy.

MBZUAI

• Built a deep learning pipeline for Africa and Pediatric brain tumor segmentation for the corresponding BraTS 2024 tasks.
• Integrated the novel schedule-free optimizer into the training algorithm.
• Performed thorough analysis of fine-tuning and ensembling techniques, and their impact on performance.

MBZUAI

• Developed a novel multimodal framework combining BioLinkBERT/CLIP embeddings with 3D Swin Transformers.
• Engineered a hybrid architecture integrating clinical text prompts and PET/CT imaging for tumor segmentation.
• Pioneered an EHR processing pipeline using semantic text embeddings.