Skip to content
View harshithkethavath's full-sized avatar
๐ŸŽฏ
Focusing
๐ŸŽฏ
Focusing

Organizations

@uga-gaim

Block or report harshithkethavath

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please donโ€™t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this userโ€™s behavior. Learn more about reporting abuse.

Report abuse
harshithkethavath/README.md

Hi there ๐Ÿ‘‹๐Ÿป

My name is Harshith Kethavath. I'm an MS Computer Science student at the University of Georgia, currently an Applied ML Engineer at Lab for Geoinformatics and AI Modeling. My current work is SignsOfExtremes, where I use computer vision for extreme weather event prediction.

๐Ÿ’ผ Open to Applied/ML Engineer and ML Systems roles.

EarthVision Workshop at CVPR 2026

Low-Data Supervised Adaptation Outperforms Prompting for Cloud Segmentation Under Domain Shift Kethavath & Hu - Proceedings of the IEEE/CVF CVPR Workshops, 2026

Zero-shot VLM prompting fails on Sentinel-2 cloud segmentation under domain shift, and the 60 engineered CLIPSeg prompts scored as low as 0.07 mIoU against a 0.255 baselne. Supervised adaptation overtakes all prompting at roughly 8 labeled images, and 5-10% of labels recover ~85% of peak mIoU.

Code, experiments and models -> uga-gaim/2026_CVPRW_CloudPrompts

Selected Work

๐Ÿ›ฐ๏ธ MarsDEMNet - Single-image elevation (DEM) prediction from Mars satellite imagery. Built and ablated 4 architectures on 80,000+ NASA image pairs; a 5-block multi-output U-Net reaches 59.9m validation RMSE via multi-task masked MAE loss over elevation, slope, and roughness. Diagnosed a Lustre filesystem bottleneck on HPC and cut per-epoch data loading from 3,800s to under 600s with RAM-cache preloading.

๐ŸŒฆ๏ธ WeatherMind - Modular VLM framework for weather analysis on sky imagery across 4 backends (Qwen, SmolVLM2, Moondream2, Gemma3) with context-aware prompting. A spatiotemporal ETL pipeline (Pandas + OCR) aligns image timestamps with ASOS station logs to generate 15,000+ ground-truth labels, alongside a zero-shot weather-feature detector on open-vocabulary object detection.

๐ŸŒŠ FloodPrint - CONUS-scale spatiotemporal verification of NWS Flash Flood Warnings for the 2024 warm season. A spatial-index + buffer-intersect join matches ~3,700 warning polygons against ~4,730 Flash Flood storm reports for a 45.2% national verification rate. Surfaced a sharp geographic gradient, 0-13% in the desert Southwest vs. 77-85% in the Northeast, and traced it to spotter-density bias rather than forecast skill.

๐Ÿ“จ Reach me at: harshith@kethavath.com

Pinned Loading

  1. uga-gaim/Pub_2026_CVPRW_CloudPrompts uga-gaim/Pub_2026_CVPRW_CloudPrompts Public

    Jupyter Notebook 4

  2. MarsDEMNet MarsDEMNet Public

    Jupyter Notebook

  3. WeatherMind WeatherMind Public

    Jupyter Notebook

  4. FloodPrint FloodPrint Public

    Jupyter Notebook