Fine-tuned a pre-trained Vision Transformer (ViT) using PyTorch on satellite imagery datasets, improving model generalization and inference performance for land
use and geospatial analysis across multiple terrain categories.
Built automated data preprocessing and feature engineering pipelines using Python to optimize geospatial and agricultural datasets for LightGBM model training,
reducing data preparation time by 30% and improving overall model reliability.
Designed a PostgreSQL database schema with PostGIS extensions to store and query geospatial datasets, structuring satellite imagery metadata and land
classification outputs for efficient retrieval by research teams.
Deployed trained models as REST APIs using FastAPI to serve real-time land classification and yield predictions, reducing average query response time to under
200ms for internal analysis tools.
Created an interactive Streamlit dashboard integrated with ElasticSearch-indexed survey reports, allowing analysts to visualize model outputs alongside historical
land-use records and cutting manual document lookup time by 50%.