Prakhar Sahu

I’m a B.Tech student focused on AI/ML, currently working as a Research Intern at the Centre for Artificial Intelligence and Robotics (CAIR), IIT Mandi. I build end-to-end systems in PyTorch, turning natural language and RGB-D signals into continuous control outputs.

In my research, I designed a multi-task architecture with a frozen CLIP text encoder and custom CNN visual encoders, jointly supervising action prediction, collision avoidance, and landmark localization. I also implemented a custom PyBullet outdoor simulation environment and an obstacle-aware oracle, producing 45k expert trajectories and scaling dataset collection to improve robust training.

Beyond robotics, I develop practical NLP solutions through projects like a semantic search engine for research papers and CLM fine-tuning using HuggingFace Transformers. I enjoy building end-to-end pipelines—from preprocessing and indexing to training configuration and evaluation—so models work reliably beyond just experiments.