Videet Mehta
I'm a computer science student at MIT working on frontier AI research, mostly multimodal LLMs and making models run faster on real hardware.
Right now I'm at HAN Lab, using diffusion language models for speculative decoding, and at MIT's Spoken Language Systems Lab making full-duplex speech models more efficient and expressive. This spring I'm joining Inco as a research intern.
Before that I was at Together AI, where I trained universal diffusion speculators using 30% of the compute. I also pretrained audio encoders for on-device deployment at Liquid AI, forecast marginal prices at Mercuria Energy Trading, and helped build India's first Hindi–English conversational speech model at Sarvam AI.
In 2024 I represented the USA at the International Olympiad in Artificial Intelligence and won a gold medal. I'm now on the organizing and scientific committee for the 2026 USA AI Olympiad team.
Writing
- Making One GPU Go Brrr: Reinforcement Learning in a Single CUDA Kernel
- Compute-Aware Hybrid Attention Architecture Search
- Fast Humanoid Control with Flow Matching
- Why PPO Beat GRPO at Teaching a Vision-Language Model to Solve Mazes
- Parallelizing CDF
Experience
- Together AI · Kernels Researcher
- Liquid AI · AI Researcher
- H.A.N Lab · AI Researcher
- Sarvam AI · Machine Learning Engineer
- Mercuria Energy Trading · AI Research Scientist Intern
- MIT CSAIL Spoken Language Systems · AI Researcher
- Hidden Studios · Founding Engineer
- Regenerative Neurotechnology Lab · AI Researcher
- Houston Learning Algorithms · AI Research Scientist