Gursimran
A selection of open-source contributions and projects in reinforcement learning, multimodal AI, and interactive visualization.
Integrated LoRA-based RL fine-tuning into AReaL's Megatron engine, including distributed adapter training and synchronization with vLLM. Supports GPU and Ascend execution, pipeline parallelism, and multi-node mixture-of-experts training.
A serving system for large multimodal models that separates encoding, prefill, and decoding into independently scalable stages. Supports intra-request parallelism and dynamic GPU resource allocation to reduce latency and improve throughput.
A framework for automatically placing visible watermarks on 3D models, balancing asset protection with usability. Includes a watermarking pipeline and support for embossing watermarks onto model surfaces.
Combined neural perception with symbolic reasoning in a Bayesian model for visual question answering. Explored inference over discrete scene representations on the sort-of-CLEVR dataset.
Explored question-specific visual filters, attention over dialog history, and end-to-end training for cooperative question-and-answer agents.
Built interactive polynomial regression and k-nearest neighbours playgrounds to show how data complexity and model choices affect bias and variance.
