Kaashyap Murali
Tech Lead at Microsoft working on digital agent co-workers for Copilot Studio (orchestration, evals, memory systems, etc.). It is going well enough that I still have the job, and strangely enough that I still have things to write about.
Currently open in my tabs: agentic search
Recent
-
Fireline: forecasting tomorrow's wildfire from satellites
Next-day wildfire spread on WildfireSpreadTS: a small U-Net, where tomorrow's fire comes from, a test of JEPA pretraining, and a leak in the benchmark's vegetation channels.
Test year 2021: AP 0.567 for next-day fire, against 0.407 for the best baseline. JEPA pretraining did not help at this data scale.
Explore the project -
DisasterLens: predicting building damage from satellite images
Per-building damage assessment from pre- and post-disaster satellite images on xBD, with Mask R-CNN building detection and a fine-tuned, satellite-pretrained DINOv3 ViT-L.
Recall on the xBD test split with known building footprints: 97% no damage, 61% minor, 73% major, 84% destroyed.
Explore the project -
(Published on Microsoft's blog) Memory in Copilot Studio
How agent memory bridges conversations in Copilot Studio, from short- and long-term memory to reflection and evaluation.
Read on Microsoft -
DeepMind's AI Co-Mathematician and the Case for Episodic Memory
DeepMind's AI Co-Mathematician treats failed proof attempts as durable project state, not noise to clear. Notes on why episodic agent memory, not the next model, is the real frontier.
Read more