Valgo uses advanced statistical simulations to figure out how likely a self-driving car is to crash on a specific route, so insurance companies can set a fair price even though no one has decades of crash data for robots yet.

Valgo's core ML use case applies Bayesian safety validation and importance sampling to estimate the failure probabilities of autonomous vehicle systems across diverse routes, environments, and operating conditions. Because physical AI systems like autonomous vehicles lack the billions of historical claims that traditional auto insurance relies on, insurers cannot use conventional actuarial methods to price coverage. Valgo addresses this by constructing high-fidelity probabilistic simulations of AV behavior, then using Bayesian inference and surrogate models to efficiently explore the tail of the risk distribution — identifying rare but catastrophic failure modes. The system leverages state-dependent importance sampling to focus computational resources on the scenarios most likely to produce losses, dramatically reducing the number of simulations needed to achieve statistically significant failure probability estimates. The output is a per-route, per-task risk score that insurers can directly integrate into their underwriting and pricing workflows, effectively creating synthetic actuarial tables for autonomous systems. This approach draws directly from founder Robert Moss's published research on Bayesian safety validation and black-box system analysis.

It's like crash-testing a million virtual self-driving cars on every road in America so the insurance company doesn't have to wait for real crashes to figure out what to charge.

Valgo mathematically proves how safe a robot's AI brain is before it gets insured, so the insurance company knows exactly what it's covering instead of just guessing.

Valgo's second key ML use case leverages neural network verification — specifically backward reachability analysis — to generate probabilistic safety certificates for the AI controllers powering insured autonomous systems. When an insurer evaluates whether to cover an autonomous vehicle, drone, or industrial robot, a critical unknown is whether the system's neural network controller will behave safely across all plausible operating conditions. Valgo addresses this by applying formal verification techniques developed by co-founder Sydney Katz, which propagate safety constraints backward through the neural network to identify the precise set of input conditions under which the system is guaranteed to remain safe. The platform then translates these verified safe operating envelopes into underwriting-friendly metrics: a confidence score reflecting the percentage of expected operating conditions that fall within the verified safe region, and a residual risk estimate for conditions outside it. This gives insurers an unprecedented level of transparency into the actual safety properties of the AI systems they are covering, enabling more precise pricing, clearer policy exclusions, and reduced reserve requirements. The approach is grounded in Katz's 20+ publications on neural network verification and AI safety.

It's like getting an X-ray of a building's foundation before writing the property insurance — except the building is a neural network and the X-ray is math.

Valgo continuously recalculates how risky a robot's current mission is as conditions change in real time, so the insurance price always matches the actual danger instead of being a stale guess from months ago.

Valgo's third novel ML use case applies constrained partially observable Markov decision process (POMDP) planning — specifically the ConstrainedZero and BetaZero algorithms developed by co-founder Robert Moss — to dynamically adjust risk estimates for insured autonomous systems during active operation. Traditional insurance pricing is static: a premium is set at policy inception and remains fixed. But autonomous systems operate in dynamic, partially observable environments where risk fluctuates dramatically — a delivery robot's risk profile changes with weather, traffic, terrain, and task complexity. Valgo's platform uses learned POMDP solvers to model the autonomous system's decision-making process under uncertainty and safety constraints, predicting how the system will behave across a distribution of future scenarios. By running these models continuously against real-time telemetry and environmental data, Valgo can provide insurers with dynamic risk scores that update as conditions evolve. This enables usage-based insurance models, real-time portfolio risk monitoring, and automated alerts when insured systems enter high-risk operating regimes. The constrained formulation ensures that safety requirements are explicitly encoded, allowing the platform to flag when an autonomous system's planned actions are likely to violate safety thresholds — a critical input for both insurers and operators.

It's like having a fitness tracker for a self-driving car that tells the insurance company in real time whether it's taking the safe scenic route or the sketchy shortcut through a construction zone.