ULTRAGENIS

ProbeLab is an AI-powered scientific experimentation platform I developed to enable artificial intelligence to go beyond answering scientific questions and actively investigate physical phenomena. Most AI systems can explain scientific concepts, but they do not directly conduct experiments, estimate unknown physical parameters from observations, quantify uncertainty, or determine which experiment should be performed next. I developed ProbeLab by combining computer vision, physics-based simulation, Bayesian inference, active experiment selection, and multimodal AI reasoning into a unified workflow. ProbeLab begins by observing a physical experiment through video. Meta's SAM 3.1 segments and tracks moving objects, converting visual observations into measurable trajectories. These measurements are processed through geometric calibration and a custom numerical physics engine that estimates unknown parameters and their uncertainty. Rather than simply analyzing existing observations, ProbeLab uses Expected Information Gain to identify which experiment could provide the most useful new information. Meta's Muse Spark interprets the scientific results and explains the reasoning behind the recommended experiment. ProbeLab can also generate predictions for unseen experimental conditions before receiving new observations. These predictions can then be compared with actual experimental results to evaluate the accuracy of the inferred physical model. For my initial prototype, I selected a nonlinear damped pendulum. The system processes pendulum videos, extracts angular motion, estimates parameters such as pendulum length and damping, and recommends informative follow-up experiments. My long-term vision is to extend ProbeLab beyond pendulums into a general-purpose AI scientific experimentalist capable of observing physical systems, evaluating hypotheses, designing experiments, and improving its understanding through evidence. ProbeLab — From Observation to Discovery.