I am a researcher and ML engineer working at the intersection of Earth system science, agriculture, climate, and production AI. I build hybrid AI-physics models that turn satellite, airborne, and ground-sensor data into crop, weather, and climate insight, and I design the production ML systems, from LLMs and VLMs to low-latency inference, that bring those models to reliable, real-world scale.
Selected work spanning production AI systems and Earth/agriculture ML. Hover to pause.
Earth Foundation Models · NeurIPS 2026 Workshop
A benchmark of AlphaEarth, OlmoEarth, Presto, and AnySat embeddings for in-season, pixel-level corn and wheat yield forecasting on YieldSAT. Tests whether foundation model embeddings capture real crop phenology over time and preserve intra-field spatial variability, compared with a multimodal fusion model trained from scratch.
LLM · RAG · Production AI
A production-ready Retrieval-Augmented Generation system for Persian finance and tax law documents. Combines hybrid BM25 + vector retrieval over a Chroma index with a streaming FastAPI backend and React frontend, plus a built-in evaluation framework with regression gates that block index deployments when answer quality drops.
Climate AI · Microsoft
A large-scale machine learning pipeline for statistical downscaling of global climate model outputs to high spatial resolution, developed at Microsoft. Covers the full ML lifecycle — from raw climate model ingestion and feature engineering to distributed training, automated validation, and deployment-ready forecast outputs.
Open Source · CVPR 2025
Open-source Python package and benchmark dataset for county-level crop yield forecasting across 70+ California crops (2008–2022). Integrates satellite imagery, climate variables, evapotranspiration, and soil layers into structured dataloaders, published to HuggingFace Hub with evaluation protocols and baseline results.
Agricultural Risk Intelligence
A weather-risk platform connecting long-range forecasts with post-event damage assessment through a dual 45-day window: pre-event heat/frost/flood risk for corn and grapes, then post-event satellite verification — cutting damage-estimate turnaround from weeks to days.
Open Source · React + Leaflet
An interactive map viewer for the California Crop Yield Benchmark dataset, built with React and Leaflet. Renders Cropland Data Layer, evapotranspiration, and gNATSGO soil layers by county and year, parsing GeoTIFF and Zarr data client-side from public cloud storage.