Vision-Language Models for Computational Neuroscience
Model adaptation and multimodal research tooling for digital histology.
Context and problem
At the Sudha Gopalakrishnan Brain Centre, my work focused on adapting language and vision-language models to digital histology and neuroscience research. The broader environment included high-resolution Nissl- and H&E-stained imagery, MRI-linked data, annotations, and biomedical metadata. Data preparation and alignment were central parts of these workflows.
The work took place around the centre’s brain-imaging efforts and the DHARANI initiative. My contribution was to model adaptation and supporting tools within that environment.
My contribution
I worked on preprocessing, image/text alignment, annotation encoding, and model adaptation. This included OpenCLIP and BLIP-2 fine-tuning and the preparation of multimodal data pipelines for research use.
I also built internal inference tools and REST API interfaces to make model execution usable within research workflows. The work connected experimentation with practical interfaces for running inference on prepared data.
Approach and implementation
Experimentation used NVIDIA DGX A100 infrastructure. The research environment involved million-scale histology imaging datasets; that describes the data context, rather than a claim that every available image or modality was jointly used in a single model.
OpenCLIP and BLIP-2 were model-adaptation components. MRI-linked material and other biomedical metadata formed part of the broader data workflows. Each workflow should be understood in its own context, rather than as one fully integrated end-to-end system.
Evidence and scope
The internship certificate confirms involvement in the Large Language Model Team from December 1, 2025 to February 28, 2026. This account describes contributions and implementation scope; it does not report a numerical accuracy improvement or clinical validation.
Research data, internal tools, and model artifacts are not reproduced here. No unpublished research images are included.