Bio and Interests
Hello and welcome to my personal website! My name is Hemil Desai, and I am a Software Engineer with a passion for developing innovative solutions in the field of Machine Learning and Artificial Intelligence. I’m intrigued by the latest developments in Generative AI, including Large Language Models like ChatGPT from OpenAI, Claude from Anthropic, open-source versions like Flan and Bloom, etc; image synthesis models like Dall-E, Stable Diffusion, etc and multi-modal research resulting in models like BLIP, CLIP, etc. I recently finetuned my dreambooth model (check out an example on the left, upscaled using SwinIR), and have even built a RTX3090 rig out of curiosity.
I earned my Master of Science from UCLA and my Bachelor of Science from Purdue University, both in Computer Science. My industrial experience includes my current gig at Snap, followed by previous stints at Roam Analytics and Braintree Payments.
Throughout my career, I have gained extensive experience in developing cutting-edge tools and infrastructure for machine learning applications. As a software engineer at Snap Inc, I led the optimization and refinement of training pipelines for various ranking use-cases, resulting in significant cost savings and improvements in engagement metrics. At Roam Analytics, I built a robust and auto-scaling GPU-enabled Kubernetes cluster on Amazon EKS. At UCLA, I worked as a researcher at the BigML Research Lab led by Professor Baharan Mirzasoleiman, where I created a distributed version of CRAIG that runs across multiple GPUs (Ask me about the most interesting problem I faced as part of this project).
I have a wide range of skills in various programming languages, frameworks, libraries, platforms, and clouds, including Python, Go, Java, PyTorch, TensorFlow, AWS, GCP, Kubernetes, and Terraform, among others.
In addition to my above experience, I have also been involved in personal projects, including the creation of an online multiplayer alternative for the popular board game Catan called Imperials which was built using Go, Websockets, and Pixijs. Please reach out at [email protected] if you want to connect.