Udi Menkes

Udi Menkes’s AI work includes shared inventorship on an Intuit-assigned method for improving language-model prompts through automated evaluation and diagnosis. AI Engineer lists him as Principal Product Manager, Intuit among past speakers and attendees. His finance-track contribution, summarized by Latent Space in July 2026, emphasized the need for financial AI to understand state, actions, outcomes and risk.[2][3][1]

In its July 29, 2026 finance-track coverage, Latent Space summarized Menkes’s Intuit contribution as arguing that generic LLMs are insufficient for financial work: models must understand real financial state, possible actions, their outcomes and associated risk. The publisher placed that requirement in the context of Intuit serving approximately 100 million consumers, small businesses and accountants.[1]

Menkes, Linoy Cohen, Matan Vetzler and Yakov Gazman are the named inventors of US patent 12,481,651 B1, “Language model improvement through automated prompt engineering”. Assigned to Intuit, the application was filed on June 30, 2025.[3]

The claimed method connects evaluation failures to prompt revision. It evaluates triplets containing a prompt, the model’s response and a pass-or-fail score against a metric. When the overall alignment score misses a threshold, a judge language model examines a failed triplet to identify the type of misalignment, then examines that diagnosis alongside the triplet to identify its cause. The judge uses the cause to modify the failed prompt and returns an enhanced prompt predicted to produce an aligned response. This is a jointly invented prompt-improvement method; the claim describes a predicted result rather than a demonstrated accuracy gain or an alignment guarantee.[3]

Before his later Intuit work, Menkes wrote about AI business strategy at OpenWeb. In a January 21, 2021 article, he proposed three tests for an “AI-Moat”: deep domain knowledge, evidence of an algorithmic edge and a first-party data strategy. He argued that domain expertise helps teams identify business problems and useful input features, describing the combination of domain expertise and data science as producing “killer features.” He recommended looking for domain-specific product flows, a data model and domain experts in leadership.[4]

Menkes expected machine learning to become increasingly commoditized and advised valuing algorithmic claims backed by recent research against industry benchmarks. For publishers, he argued that data collected directly from community participation could inform audience experiences, advertising and editorial choices. These were strategic arguments about potential advantages, rather than measured results. His article also described recruiting a data engineering team lead, product data analysts and data scientists for OpenWeb.[4]