Sajil Awale

Graduate Research Assistant @Nasa-Impact

US

I develop advanced NLP and agentic AI solutions for Earth science data—focusing on keyword classification, domain-specific language models, workflow automation, and foundation models for scientific discovery. Science Keyword Recommender – Built an extreme multi-label classification system for NASA’s CMR, expanding the science keyword set from 430 → 3,240 and improving data discoverability. Achieved F1 = 0.55 using Focal Loss and stratified dataset splitting.🧠 Pretraining Science Document Embeddings – Pretrained Indus-SDE, a RoBERTa-based model, on 520k NASA SDE documents (1024-token context, Weighted Dynamic Masking), achieving 78.1% top-1 MLM accuracy and outperforming baselines on downstream Earth science and astrophysics tasks. Downstream Unification Framework – Developed a multi-task fine-tuning and evaluation framework (Hugging Face + W&B) with YAML-driven pipelines, unified preprocessing, automated analytics, and experiment tracking—standardizing and accelerating NLP model development. Indus-SDE-ST Sentence Transformer – Currently training a multi-GPU Sentence Transformer on large-scale scientific text/code pairs for retrieval and reasoning. Early results show state-of-the-art embedding performance, enabling faster scientific discovery.🤖 Agentic AI Evaluation & Benchmarking – Conducted comparative evaluations of NASA’s Deep Literature Search Agent against Gemini and OpenAI systems using LLM-as-judge metrics (contextual precision, recall, relevance, faithfulness). Explored agent reliability, metric stability, and variance reduction strategies to improve reproducibility and trust in autonomous scientific research agents.

Pulchowk Campus

Bachelor's degree, Electrical, Electronics and Communications Engineering

The University of Alabama in Huntsville

Masters, Computer Science