Isaac Emevo

I developed a financial dataset and benchmarking framework to evaluate LLMs across STEM quantitative computation and generalist domain reasoning. For the STEM tier, I curated multi-step mathematical workflows—including derivative pricing, portfolio optimization, time-series forecasting, and code-based modeling—to assess numerical precision and algorithmic soundness. For the generalist tier, I designed tasks covering 10-K/10-Q synthesis, earnings call sentiment analysis, regulatory compliance, and advisory dialogues to evaluate macro-level financial literacy and strategic context.
I implemented a dual-pipeline evaluation methodology to ensure data fidelity. Quantitative outputs were validated via deterministic programmatic verification against exact ground-truth solutions under strict error tolerances. Qualitative, reasoning-heavy responses were assessed using structured rubrics combining automated LLM-as-a-judge scoring with expert human review for factual consistency. By bridging computational finance with commercial reasoning, this benchmark reduced hallucinations, measured domain proficiency, and accelerated reliable financial AI adoption.

The goal of the project was to develop domain-specific datasets to train advanced Large Language Models (LLMs) to effectively interact with visual content from specific domains. I was a creator and reviewer on this initiative to analyze complex textbook images, charts, and diagrams to formulate non-extractive prompts requiring high-level cognitive effort and visual interpretation. I created detailed explanations with a strict “Observe, Explain, Answer” framework to trace out logical, line-by-line deductions that demonstrate expert domain knowledge. Additionally, I conducted extensive quality validation audits on visual question answering (VQA) tasks, evaluating for scientific accuracy, coherence, complexity, and strict format compliance to maximize data integrity for AI training.