About

prasanna-kotyal

Hi, I'm Prasanna.

I'm doing my MS at IIIT Bangalore, and I work at a defense tech organization on distributed pre-training and inference optimization of AI models. Mostly vision models, running on edge devices.

Before this I verified and validated vision models, generated domain-specific synthetic data for pre-training and fine-tuning at scale, and built our best performing regression model on Jetson devices.

I made this site to write mini essays. Some will be debugging notes from GPU work. Others will be about whatever I've been chewing on lately: a paper, a tool, a book, an opinion I want to pressure-test. Writing it down is how I find out whether I actually understand a thing or just recognize it.

The design of this site is stolen from ben-mini.com , with the greens turned blue. Stealing is a skill.

Links

• LinkedIn
• X - @prasannakotyal
• GitHub - prasannakotyal
• Email

Projects

• 2026
  • Mates: A replayable scrapbook of my online checkmates, with just enough of the setup to make each finish worth watching. (mates.prasannakotyal.com)
  • Distributed Transformer: Trains a decoder-only Transformer with DDP and FSDP and measures when each is worth it. (GitHub)
• 2025
  • Flash Attention (CUDA): Flash Attention in raw CUDA, checked against a naive kernel and a CPU reference. (GitHub)
  • Pokedex: Quick-lookup pokedex, built to beat a friend at Pokemon Showdown. (GitHub)
• 2023
  • Real-Time Hand Gesture Recognition: Recognizes a set of hand signs live from a webcam. (GitHub)