projects

GPU MODE Competitions

• NVIDIA NVFP4 Blackwell competition (2026)
• AMD x GPU MODE E2E Model Speedrun — $1.1M prize pool, MXFP4 MoE / MLA Decode / MXFP4 GEMM kernels and end-to-end DeepSeek-R1 / Kimi K2.5 inference on MI355X (2026)
• AMD Developer Challenge 2025: Distributed Inference Kernels — $100K grand prize, multi-GPU All-to-All / GEMM+ReduceScatter / AllGather+GEMM kernels on MI300X (2025)
• AMD Developer Challenge 2025: Inference Sprint — $100K grand prize, MLA with RoPE / Fused MoE / FP8 GEMM kernels (2025)
• Jane Street model optimization competition (2025)

AI/ML

• Efficient Triton Implementation of MSA pair weighted averaging with gating for AlphaFold3 🧬
• Triton Implementation of FlashAttention2 with Custom Attention Masks 🎭
• A Simple Framework for Intrinsic Reward-Shaping for Reinforcement Learning using LLM Feedback (bad code)
• Pruning Methods with Sparse Factorizations of Neural Networks
• Appearance Codes using Joint Embedding Learning of Multiple Modalities
• Predicting Visual Futures with Image Captioning and Pre-Trained Language Models
• Developing a Device-Independent Software for Adaptive Tremor Suppression Orthoses

  - Intel International Science and Engineering Fair (ISEF) 2019 - 2nd Place

Math Stuff

• Robustness of Overparameterized Deep Learning
• Transaction Fee Mining and Mechanism Design
• Quantum Information Theory and the No-Cloning Theorem