Sababa Saad Usmani

Applied ML/NLP to analyze real-world investor–founder communication for empirical research. Built end-to-end pipelines to transform unstructured transcripts into structured data via cleaning, normalization, and multidimensional feature engineering. Used transformer embeddings and classical ML to extract behavioral signals like sentiment, conviction, and uncertainty. Developed predictive models on imbalanced data, prioritizing interpretability to link communication patterns to investment outcomes.

Developed a 1.0-grade Master’s thesis on hybrid ML-LLM document classification for imbalanced data. Built a three-layer pipeline (XGBoost/TF-IDF, Transformer uncertainty detection, and GPT-4o RAG) that improved macro-F1. Identified the "anti-breakage" principle to prevent LLM corrections from degrading accuracy, turning a potential 66% drop into a performance gain. Validated on 20 Newsgroups (0.88 F1) and deployed via Gradio. Research manuscript currently under review at ICDAR 2026.

Led ML systems for classification and automation in production. I built scalable pipelines for structured and unstructured data, applying feature engineering and model optimization to boost performance. I deployed robust models for domain-specific problems, managing noisy datasets and data quality. By coordinating cross-functional teams, I translated business needs into maintainable, high-scale architectures, ensuring long-term reliability and effective decision-support solutions.