Santosh Sapkota

I build production-focused machine learning systems that turn messy documents into usable data, from preprocessing to OCR fallbacks. At Ishani Technology, I designed a hybrid extraction pipeline that reached near 100% accuracy on digital PDFs and ~85% on scanned municipal documents, then extended it to attendance sheets with CNN-based handwriting classification (~90%). I also integrated an LLM-based multilingual translation pipeline with parallel processing for efficient multi-page document conversion.

Previously, as a Machine Learning Intern at Elevvo Tech, I developed end-to-end ML pipelines across 4 projects, using SMOTE for class imbalance and rigorous feature engineering. I’m especially drawn to MLOps-style orchestration—building an ICU early-warning system operationalized with MLflow, DVC, and Airflow, and deploying containerized FastAPI services with monitoring for reliability.