Abhishek Nair

I work on GPU-accelerated machine learning systems, focusing on everything from kernel optimization to model training and deployment. I’m interested in exploring new approaches and building practical solutions that push the boundaries of what’s computationally feasible.​​​​​​​​​​​​​​​​

I like structure, minimalism, abstraction (the irony keeps me accountable), and learning more about the unknown.

I also like top-rope climbing and scrambling, currently climbing 5.10a (YDS).

Projects

• occupancy.nn - Realtime Scene Reconstruction (Ongoing)

  A unified platform to train occupancy networks with distributed computing and run accelerated inference to reconstruct 3D scenes in realtime

• llama3.cu - An LLM Inference Engine

  A CUDA native implementation of the Llama 3 architecture for causal language modeling using custom CUDA kernels for accelerated inference.

• nerf.3d - 3D Scene Reconstruction with NeRFs

  A PyTorch implementation of the NeRF: Representing Scenes as Neural Radiance Field for View Synthesis paper by Ben Mildenhall, et al.