Arva AI opens Research Lab to automate high-risk fraud decisions

Arva AI launched a Research Lab to develop proprietary models and an operational layer intended to automate the highest-risk decisions in financial crime and fraud. The company reported its first model, Arva Intel, delivered 13% higher precision than general-purpose models in independent testing.

Backed by Y Combinator and Google’s AI Fund, the Lab logged more than 5,000 hours of research, training and evaluation. The effort has produced two outputs already running in production at banks: specialized models that handle the riskiest components of investigations, and an infrastructure layer called AgentCore that converts analyst corrections, insights and case outcomes into system updates.

Arva Intel currently focuses on enrichment tasks, such as researching suspicious individuals and businesses online. The company plans to expand model capabilities to transaction analysis and evidence-based reasoning and to apply the Lab’s work to payment exceptions, disputes and other customer-related investigations.

The company said its benchmarks measure the accuracy of each component inside a case-for example the correctness of enrichment results or the identification of relevant evidence-rather than only the final case outcome. In the independent evaluation cited by Arva, Arva Intel scored 13% higher precision than frontier general-purpose models on those component-level measures.

AgentCore includes backtesting, evaluation and version control before any model update reaches live decisioning, the company reported. Arva also reported the technology is in use at global financial institutions, including a top-10 U.S. bank.

Rhim Shah, Arva AI’s founder and CEO, described the industry challenge: ‘Banks keep humans in the loop because no AI has been accurate enough to remove them safely.’ He described the Lab’s models and AgentCore as tools to automate certain decisions while maintaining the accuracy and controls banks require.

The company plans to publish its benchmark methodology and research at academic venues. Arva described its approach as combining domain-specific models with an operational layer that continuously incorporates analyst feedback and maintains traceability and versioning for model updates.