Hi, I'm Sagnik Bhattacharjee.
I'm a researcher working on agent reinforcement learning and harnesses for long horizon tasks.
Currently, I'm interning at a neolab where we're trying to accelerate scientific discovery. Previously, I worked at the Language Technologies Research Centre on DPPNets and optimized kernels for speech models.
Some of my featured work:
- Post training a model to explore demoscenes
- Reverse engineering the Qualcomm NPU compiler
- Sampling NCA trajectories to pretrain SLMs faster
- NPUs: What they are and where they break
- Writing a ~1200 Elo chess engine under 2KB
More of my research and writings can be found here.
Outside of work, I love playing around with computers and making art with them. I've spent a good part of my life making music, painting and exploring literature, cinema and conlangs. I also help run Threadlocked, a community trying to bring back the spirit of the old hacker culture in India & APAC.