• Co-authored a research paper proposing LaGPS, a neuro-symbolic framework for detecting human-centric rule violations from images.

• Introduced the HRS benchmark (1,100 annotated images) and achieved +19.4% mIoU improvement over CLIPSeg and Grounded SAM baselines.

• Researching animal 3D reconstruction using SMAL + SDFit, diffusion models and ControlNet conditioning. Working on pose alignment, mesh fitting stability and reconstruction quality.

• Worked on satellite-based landslide detection using multi-spectral imagery with YOLO, UNet and Transformer segmentation models.