Part of a pre-seed startup focused on personalized LLM applications. Led core AI initiatives including fine-tuning foundation models, building RAG pipelines, and deploying scalable LLM APIs. Collaborated closely with founding team to shape the architecture and build the prototype.

• Conducted a literature survey on the existing time-series graph-based ML frameworks - TGN, TGL, NeutronStream

• 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.

• Built baseline models for VWSD using InternVL2 VLM, and context augmentation using LLMs and lexical databases.

• 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

• Built a keypoint detection model using YOLOv8 for automatic keypoint annotation of drone whale images

• 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%

• Developed a multi-agent based long-form article generator using LangChain and RAG-based expert simulations for content generation.

• 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.