Perle AI

Perle AI operates a human-expert data layer for organizations building and deploying AI models. Its offering is organized around three product lines: managed programs that run from brief to delivery covering training data, evaluation and labeling on a unified platform; an expert-sourcing capability that lets customers scale teams with vetted specialists using AI-assisted interviews, domain rubrics and quality signals; and continuously refreshed, expert-labeled multimodal datasets for teams building real-world robotics systems.

The company focuses on problems that arise once readily available data has been exhausted and models still fail on difficult cases. Work includes stress-testing models beyond standard benchmarks using an evaluation network of physicians, attorneys and specialists across roughly two dozen disciplines who author rubrics, run cases and document failures; and expert-led red-teaming, refusal calibration and bias auditing intended to identify where a model breaks in terms a safety team can act on. For robotics, its WhisperMind effort captures human manipulation in everyday environments — such as picking up a wet glass, opening a stuck drawer or locating an item in an unfamiliar kitchen — as multi-sensor RGB-D, IMU and force-torque streams labeled by people with direct task experience.

Perle also publishes benchmark research on frontier model behavior, including work on whether foundation models can identify sound sources, leniency of frontier models when grading Arabic cultural knowledge, and updated automatic speech recognition code-switching benchmarks for Arabic, Persian and German. Prospective customers can request evaluation of their own models or languages against these benchmarks.

Business model

Perle delivers expert human data and evaluation work as managed programs run end-to-end by its own expert teams on a unified platform, and supplies access to vetted specialists and to recurring expert-labeled multimodal datasets. All experts in the network are reported to be on payroll rather than engaged as an ad hoc crowd.

Traction

The company reports a network of more than 15,000 vetted experts across 70 countries and 76 languages, including 2,500 physicians and 800 attorneys plus specialists across about two dozen disciplines, and states it works with hospital systems, law firms, robotics labs and AI model teams.