Sandesh Pokhrel

I’m an early-career Software Engineer (R&D) focused on building AI-driven prototypes that deliver measurable outcomes. I thrive on system-level solutions—turning models into production-ready ML workflows with clean ingestion, preprocessing, training, and API integration.

In my current role, I built and integrated multi-step FMCG demand forecasting and RAG-based compliance evaluation systems. Using AWS Bedrock Nova Pro LLMs and an OpenSearch vector store, I processed 100–500 pages per assessment and achieved 95% top-5 retrieval accuracy.

I also engineered forecasting features—lag/rolling stats, promotions, seasonality, and external signals—improving WMAPE by 38% versus baseline. For the forecasting engine, I delivered a multi-step system predicting SKU-level daily demand across 50 SKUs × 20 locations, reaching 38% WMAPE and a 25% improvement over baseline statistical models.

I’ve complemented this work with an internship where I cleaned and prepared structured datasets, ran EDA and feature engineering for decision-making, and improved reliability through better logging and documentation for repeatable pipelines. I collaborate closely with senior engineers to iterate on architecture, troubleshoot models, and produce production-quality documentation.