At NucleusAI, I contribute to fine-tuning and data compression pipelines for vision-language models. I researched frame reduction that cut the number of frames from ~2,800 to ~400.
At Bharat Law, I built a hybrid LLM memory system using BM25 retrieval and custom mathematical ranking. I also built multi-day litigation-database crawlers, a Tauri desktop app, and an AWS SES/SQS email service.
At San Jose State University, I contributed to system design for ErisML/DEME, focused on ethics-aware governance architectures for AI agents. I also assisted with research and evaluation of Prometheus.
At Clairvyn, I developed a pipeline for GSDIFF research in floorplan generation and hand-coded an edge transformer that removed unnecessary edges. In my projects, I implemented generative models and core research architectures, including a Stable Diffusion model and a transformer built from scratch.
