Creating Human-like AI Avatars

Research on multimodal LLMs• Pioneered creation of V-Zen, a groundbreaking Multimodal LLM, revolutionizing GUI understanding and automation• Engineered a high-precision grounding module and a dual-resolution image encoder, enabling accurate GUI interactions• Leveraged DeepSpeed framework for optimal model training on an 8 NVIDIA’s A100 platform for multi-GPU training

1+ years of R&D experience in the development of innovative AI solutions for various applications• Deployed SOTA object detection & classification models for real-time AI camera product on Jetson Nano using TensorRT• Developed a scalable distributed system using gRPC & Docker, ensuring seamless deployment management & deployment• Hands-on experience in implementing training & inference pipelines for advanced NLP models across various applications• Boosted accuracy by 42% through unsupervised pre-training of Wav2Vec2 on a large Japanese speech2text dataset

Led pioneering research on OCR at the Computer Vision Lab proposing a novel approach, outperforming all the SOTA models on benchmark datasets- First-author paper Accepted at ICDAR 2023 Conference (the world’s most prestigious conference in the field of document analysis)- Project Website - https://abdur75648.github.io/UTRNet/

Worked collaboratively with a chemical engineer on predictive modelling of iron ore reduction with different fuels