in progress
Research
Open questions I'm digging into — LLM agents and remote-sensing for climate.
The Serialization Bottleneck: How Text-Based Intermediate Representations Degrade Compositional Tool-Use Agents
Investigating how forcing tool-use LLM agents to serialize intermediate state through natural language degrades compositional reasoning. Studying failure modes where text-based IRs lose structural fidelity across multi-step tool chains, and exploring structured-IR alternatives.
Climate-Driven Shifts in Teesta River Flow and Their Impact on Northern Bangladesh Agriculture: A Satellite-Imagery and Machine Learning Approach
Quantifying how climate change is altering the Teesta river's seasonal water flow and tracing the downstream effect on crop patterns and agricultural productivity across northern Bangladesh. Combines multi-year satellite imagery with machine learning models to detect flow anomalies and correlate them with regional agricultural outcomes.