Niloy R

Research Assistant

ResearchCollab • Meio período

Sep 2025 - Present • 9 mos

Developed a multimodal AI evaluation framework for Video Question Answering models using PyTorch, Hugging Face Transformers, OpenCV, CUDA, and Python. ○ Built reusable model-adapter pipelines for Qwen-VL, LLaVA, SmolVLM, InternVL, X-CLIP, and related vision language models, standardizing inference, confidence scoring, metadata logging, and result comparison. ○ Implemented attention visualization workflows to analyze model reasoning across video frames using patch-level heatmaps, frame overlays, and spatio-temporal attention summaries. ○ Designed model evaluation metrics including spatial IoU, Dice/F1, AUPRC, AUROC, temporal IoU, top-k frame recall, and deletion/insertion-based faithfulness testing. ○ Automated batch experiments, generated JSON/CSV evaluation reports, and prepared reproducible outputs for research review, technical documentation, and publication support.