Ajah

Ajah is a self-hostable gateway designed to mediate interactions between your application and various large language model (LLM) providers. It provides a robust set of features to enhance reliability and observability: • Verifies RAG responses against source documents • Flags hallucinations with zero added latency • Traces multi-agent sessions visually • Masks PII before storage • Attributes costs by user, feature, or agent step This proxy can route requests to nine different providers, including OpenAI, Anthropic, and Grok, all through a single base URL. It operates by asynchronously scoring every response for hallucination risk, RAG grounding, and PII detection, ensuring no added latency to the primary request path. The system also includes cross-model verification for an additional layer of reliability. Ajah helps engineering teams and developers who need to ensure the quality and cost-effectiveness of their language model integrations. It's an open-source solution, licensed under MIT, that can be set up quickly with Docker Compose. All data remains on your server, offering enhanced privacy and control. The platform includes a dashboard for real-time observation of traces, warnings, and costs, with configurable webhook alerts for budget thresholds and quality issues.