Arcanum

Arcanum is an applied research lab focused on evaluation and model systems for complex enterprise AI agents. It works with organizations to benchmark frontier models against the demands of a specific use case, then develops a tailored system designed to perform better with lower latency and inference costs. The company’s forward-deployed researchers and engineers collaborate with client teams to create realistic evaluation suites that reflect production environments.

Based on those benchmarks, Arcanum builds and continually improves model systems combining trained models, agent harnesses, and model routing. Its offering includes proprietary trained language models, agent verification environments, and continual learning loops that turn production failures and user corrections into new evaluation cases and training data. For challenging or low-confidence requests, systems can route work to a frontier model while handling other requests with the trained model.

Clients can own the trained model weights and run them on their own infrastructure, or choose hosted inference. Arcanum says it deploys its systems within some of the world’s largest and most important organizations