Vysakh Ramakrishnan

Designed and deployed an LLM-backed semantic retrieval system enabling temporal context awareness, efficient resource reuse, and telemetry to accelerate iteration cycles, achieving a 50% reduction in latency. Implemented advanced query interpretation and contextual filtering with unit-tested support for contextual roles, and extended data indexing with normalized metadata to ensure consistent cross-channel searchability. Collaborated with ML and backend teams to orchestrate end-to-end evaluation and deployment in a containerized environment.

Developing a temporally-aware semantic retrieval system powered by large language models, with adaptive query understanding and dynamic context tracking for improved multi-session information grounding. Built a differentiable framework that jointly optimizes physical metasurface structures and neural image reconstruction to improve imaging performance.

Formulated a minimally supervised segmentation technique enabling real-time object tracking. Created a Visual Question Localized-Answering (VQLA) module to recognize key anatomical features in surgery. Achieved robotic arm segmentation on an in-house thoracic surgery dataset using a UNet model with an IoU value of more than 92% on the validation set. Performed multi-class segmentation of the Pulmonary artery using a Mask R-CNN model with a ResNet-50 backbone pre-trained on ImageNet, obtaining IoU in the range of 81–89%.