Prankur Rusia — CTO

Working at the intersection of reinforcement learning and large language models, current focus is on post-training: the process of shaping model behavior after pretraining through RL, reward modeling, and evaluation. Currently at Microsoft SuperIntelligence, I design and run RL-based post-training pipelines for GPT models with responsibilities spanning dataset curation, RL reward pipelines, evaluation design, and training at scale. My research interests include reasoning control in LLMs, knowledge distillation via RL, and agentic systems with structured rollout design. Some problems I've worked on: • Controlling chain-of-thought behavior in reasoning models using pure RL, converting between reasoning and chat modes without significant capability loss • Sparse RL-based KD for efficient model compression • Structured agentic harnesses for long horizon LLMs • Reward design for real-world tasks where ground truth is noisy or implicit Before this, I spent time at Qualcomm on edge-deployed vision models and at ISRO as a scientist, where I published work on deep learning for atmospheric correction of satellite imagery. I'm a Kaggle Competitions Expert (top 0.6%, rank ~1100 of 180k+), with multiple medals across NLP, vision, and time-series competitions. I'm broadly interested in the science of making models more capable, controllable, and efficient through post-training and in the open questions that sit at the boundary of RL theory and large-scale empirical practice.

Stackforce AI infers this person is a leading expert in AI and machine learning, specializing in reinforcement learning and large language models.

Location: Bengaluru, Karnataka, India

Experience: 11 yrs 6 mos