Scientific research is often bottlenecked by the time-consuming creation of custom software for hypothesis evaluation. This paper introduces an AI system built with Gemini that generates expert-level empirical software for this purpose. The system takes a defined problem and evaluation method as input, proposing novel concepts and implementing them as code. It then iterates through thousands of code variants to optimize performance using a tree search strategy. The system was tested on six multidisciplinary benchmarks, achieving expert-level results across genomics, public health, geospatial analysis, neuroscience, time-series forecasting, and numerical analysis. Empirical software is designed to maximize a predefined quality score, and scorable tasks are those addressable by this type of software. The AI system generates research ideas, implements them as executable code, and uses an LLM to refine the code for improved scores. This process significantly reduces exploration time from months to hours or days, producing verifiable, interpretable, and reproducible solutions. The AI system demonstrated proficiency by generating novel solutions to challenging problems, outperforming existing expert-developed methods in several benchmarks, including predicting COVID-19 hospitalizations and integrating single-cell RNA sequencing data. This advancement promises to accelerate scientific discovery by allowing researchers to explore a vast number of potential solutions rapidly.