Kareem-Rashed/rubric-eval

Provides first-class agent evaluation beyond final outputs—assessing tool calls, execution traces, latency, and task completion—while offering zero required dependencies and native pytest integration as a neutral, MIT-licensed alternative to company-owned frameworks. Supports any LLM as a callable judge (OpenAI, Anthropic, Ollama, local) and includes optional metrics for semantic similarity, ROUGE scoring, and cost tracking, with results exportable to local HTML dashboards or JSON for CI/CD pipelines.

Available on PyPI.