Agentic Multi-Modal Triage of Citizen Flood Reports at Scale

Constructed and released the Close-View Flood Dataset (CVFD) on Kaggle: 7,855 images with pixel-wise semantic segmentation masks across 3 operationally motivated rescue classes, with inter-annotator agreement of Cohen's κ = 0.81. Benchmarked 7 vision architectures on CVFD and identified fine-tuned DINOv3 as optimal at 98.84% accuracy (±0.08) with 21.6M parameters, matching Swin Transformer at one-quarter the parameter count.

Built an agentic triage pipeline fusing a CV Agent (DINOv3 + SAM 3) with a zero-shot multilingual LM Agent into a composite priority scorer, validated via Spearman's ρ and Cohen's κ.