Developing a memory-state and dependency-graph based graph update mechanism to enable efficient and scalable training of graph neural networks on GPUs.

Implemented and benchmarked NeutronStream (Chaoyi Chen et al 2023) and ETC (Gao et al 2024) algorithms in DGL and PyG frameworks comparing link prediction accuracies and execution time with larger batch sizes.

Experimented fine-grained image ranking with semantic similarity, Tree-of-Thought and knowledge graphs.

Developed a novel re-ranking method for fine-grained image-text matching using in-context learning, attribute decomposition and multi-step prompting.

Achieved an 8.7 percent points increase in top-1 accuracy for zero-shot matching and 0.8 points increase in MRR over fine-tuned SOTA

Implemented and tested heatmap-based keypoint detection models using UNet and ResNet.

Used active learning to fine-tune the model detections and reduce the initial annotation time by 80%

Integrated FAISS for retrieval and implemented tool-calling pipelines to allow agents to access external documents and APIs.

Designed planner, writer and editor agents with feedback loops using tool calling and retrieval orchestration.

Developed a custom LLM-as-a-judge evaluation framework, achieving a 70% win rate over GPT-4o and a 5.04-point QA metric gain.