Ashr

Ashr builds custom models for organizations that have proprietary data and expert corrections that general-purpose models do not reflect. The service captures corrections made by domain experts, defines correctness through schemas and rubrics, and trains models against a grader built to that definition, all within the customer's cloud environment.

The workflow covers four stages: capturing training-ready data from documents, transcripts, logs, and team corrections; fine-tuning and reinforcement learning against a grader; benchmarking the resulting model head-to-head against the customer's current frontier API on a frozen set of their own data; and ongoing retraining as new corrections arrive. Ashr also offers an evals platform for running agents through testing, along with SDKs and API documentation for Python and TypeScript integration.

Ashr is aimed at teams that want their organizational judgment encoded into model weights rather than prompts, with the weights remaining the customer's property. The company works directly in the customer's environment, and pricing is discussed through a call with the team.