Adhitya Mohan

MS student in Computer Science at CU Boulder. I work in the Boulder AI and Systems Infrastructure Lab (BASIL) with Prof. Mark Zhao on computer systems for machine learning, and with the Human Interaction and Robotics Group on contact-aware robot control.

My work sits at the edge and on robots: getting large transformer models to run on NPUs and GPUs that were never meant to hold them, and running GPU-parallel robot simulation for sim-to-real RL. I reverse-engineered the Rockchip RK3588 NPU to deploy a vision-language model that the vendor toolchain could not compile (SHARD, EuroMLSys’26), built an open compiler and runtime for that NPU, and ported all 130 LIBERO manipulation tasks to MuJoCo Warp, where one $700 GPU runs thousands of parallel robot environments (writeup, code).

Over the summer I interned on AMD’s Research and Advanced Development team, working on sim-to-real RL: training pick-and-lift policies with MuJoCo Warp on Instinct and Radeon GPUs, then deploying them on a real XArm-6 with a classical perception front-end. That work became a co-authored ROCm blog post. Along the way I upstreamed ROCm and RDNA4 fixes to PyTorch and AMD Research’s Ryzers, including ROCm packaging for Drake and GPU-accelerated robotics stacks on Strix Point.

Before CU Boulder I was an embedded software developer at Trinamix (BASF), doing Rockchip board bring-up and V4L2 camera drivers, and an intern at NXP Semiconductors on automotive radar calibration.

I’m looking for a role starting May 2027. If you’re hiring for ML systems, edge/GPU deployment, or robotics infrastructure, get in touch.

Outside of work, I enjoy exploring geopolitics and occasionally write about topics ranging from technical deep-dives to broader socio-political analyses.