Gunjan Paul

Built multi-agent knowledge graph and RAG systems for continuous change detection, fact validation, and contradiction detection across technology and business workflows. Architected and deployed fine-tuned Llama-based models for climate and synthetic biotechnology research, improving analysis quality. Implemented vLLM optimization reducing inference latency and increasing throughput. Designed scalable multi-agent LLM infrastructure with load balancing and dynamic resource allocation handling 20K+ daily requests. Created automated graph generation for production flow visualization and built an LLM-driven grant recommendation system to match projects with opportunities.

Working on multi-modal 2D to 3D Gaussian Splatting with prompting and hierarchical segmentation in collaboration with CVLab EPFL. Implemented graph-based dense and overlapped object counting using 3D Gaussian Splat. Generated synthetic 3D semantic/instance segmentation data for plant-related CV research using Blender. Developed robust registration for low-overlap point clouds.

Developing multi-modal 2D-to-3D Gaussian Splatting with language prompting and hierarchical segmentation, in collaboration with CVLab EPFL. Graph-based dense and overlapped object counting using 3D Gaussian Splat. Created synthetic 3D semantic and instance segmentation data using Blender for plant-related computer vision research. Implemented robust registration for low-overlap point clouds.