Developed a comprehensive 16M+ image dataset with rich synthetic captions, enhancing multimodal AI training through rigorous quality control and ethical safeguards while enabling significant improvements in text understanding for vision-language models.

Fine-tuning Stable Diffusion 3 Medium at 1M-image scale with extended context window capabilities, employing model and data sharding, latent caching, and distributed training optimizations to target improved text rendering in generated images.

A distributed training implementation for class-conditional image synthesis using LlamaGen architecture with the Cosmos DI8-8 tokenizer. This framework enables efficient multi-node training on pre-cached ImageNet latents, designed for exploring autoregressive image generation model performance at scale.

Implemented PTQ techniques like logarithmic and per-tensor for the Cosmos tokenizer, showing logarithmic quantization's superior performance (25.15 dB vs 14.84 dB PSNR at 6-bit), while achieving up to 7.8× model size reduction through asymmetric encoder-decoder compression without sacrificing image fidelity.

Other Projects

At the R&D division of a biotech startup, I develop ML frameworks for Proteogenomics and Structural Genomics data analysis. I work on diverse ML research problems associated with novel Open Reading Frames (nORFs): Structure prediction, protein clustering, subcellular localization, variant prioritization, protein-protein interactions, and the evolution of novel proteins. Under Dr. Matt Wayland, I have contributed to developing our enterprise's cloud and data infrastructure.

No Links (the proprietary algorithms, pipelines, and research ideas are confidential as per CDA.)

Worked on classification, semantic segmentation, instance segmentation, clustering, and artificial image synthesis tasks for endoscopic abnormalities present in Wireless Capsule Endoscopy (WCE) images. Our team has developed an end-to-end WCE system that reduces screening time of WCE lesions from ~4 hours to ~8 minutes. We worked in collaboration with The Command Hospital Air Force Bangalore and Kasturba Medical College, Manipal.

Presentation / Paper / Video / Grant / Resources

A novel, noise-robust and computationally efficient algorithm for real-time bilateral edge detection. By integrating the concept of steerability into the design, we achieve a favorable balance between image accuracy and processing speed. Our approach builds upon Sanjay Ghosh et al.'s work on "Fast Bilateral Filtering Using Fourier Kernels," extending it with directional sensitivity for enhanced edge detection capabilities across various noise conditions while maintaining computational efficiency.

Presentation / Paper / Report

Developed a specialized pre-processing algorithm for raw textual scene images that optimizes input for OCR systems like Tesseract. Enhanced Tesseract's capabilities by training it on diverse Kannada font combinations, significantly improving recognition of Kannada script. Designed a high-performing Convolutional Neural Network for Kannada character recognition that achieved 96.8% accuracy—surpassing state-of-the-art results by approximately 2%. Implemented these advances in a practical Android application that efficiently pre-processes images and transliterates Kannada text using Tesseract tools for mobile platforms.

Report / Presentation / Pre-processing Code / Poster

Updates

Teaching Assistant for CMSC 250: Discrete Structures at University of Maryland, College Park (Spring'25)