Distributional is a free, open, and installable analytics platform for production AI agents. It helps teams understand their AI products by converting raw trace data into actionable behavioral insights. The platform enriches production AI logs and traces with statistical metrics, attributes, evals, and LLM-as-judge metrics to create a high-fidelity representation of behavioral state. It then runs continuous, adaptive unsupervised analysis including high-dimensional clustering, topic modeling, anomaly detection, and change detection to uncover behavioral signals. These insights are published as human-readable notifications on any deviations from thresholds for tracked metrics.
Key Features
- Trace-to-Insight Pipeline: Augment logs with metrics and evals, analyze for behavioral signals, and publish relevant insights.
- Enterprise Controls: Deploy to your VPC with options for local or scalable Kubernetes clusters, with authentication, access permissions, networking, and privacy.
- Flexible Integrations: Ingest logs via OTEL, SQL, or the Distributional SDK. Bring your own evals and metrics, and use existing LLM providers and frameworks.
- Continuous Monitoring: Stay current with daily insights uncovering clusters, changes, outliers, and trends in inputs, responses, tools, and context.
Use Cases
- Discover behavioral changes or opportunities for improvement in production agents.
- Identify outliers in cost, quality, or speed of AI agent data sequences.
- Triage insights rapidly with contextual evidence from relevant traces.
Key Benefits
- Converts production AI logs into actionable behavioral signals
- Continuously runs adaptive unsupervised analysis including clustering and anomaly detection
- Provides enterprise-level security controls with flexible deployment options
- Integrates via OTEL, SQL, and SDK with support for custom evals and metrics
