ajah
Self-hosted LLM observability platform providing gateway proxy, cost attribution, PII masking, quality scoring, and alerting via Docker Compose.
Ficha actualizada 23 sept 2026, 8:33 · upstream 23 sept 2026, 8:09 · release sin release detectada
Qué es ajah
Descripción editorial construida a partir del repositorio, su documentación y sus señales de mantenimiento.
A comprehensive self-hostable LLM observability solution that intercepts API calls to any supported provider (OpenAI, Anthropic, Groq, Gemini, etc.), attributes costs per user and feature, masks sensitive data like emails and credit cards, scores response quality using local ML models, and flags hallucinations. Built with Go for the gateway and Python for scoring, it runs entirely on-premise with no data leaving your network. The dashboard provides live traces, session explorers, cost charts, warning feeds, and alert history.
- Monitor LLM costs per user to identify expensive usage patterns
- Detect and mask PII (emails, phone numbers, credit cards) before storage
- Score response quality with hallucination detection using local models
- Alert on cost spikes or unusual request volumes
- Track agent session flows for debugging complex workflows
- Verify RAG responses against source documents to detect hallucinations
- Single Docker Compose deployment with zero configuration required
- Auto-detects LLM providers from API key prefixes without manual routing
- Runs quality scoring and PII detection locally on-premise with no data exfiltration
- Provides real-time cost attribution per feature and user for budget management
- Supports RAG verification with source document grounding validation
- Quality scoring model requires local hardware resources (CPU/GPU) to run sentence transformers and toxicBERT
- ClickHouse may struggle under extremely high write throughput without proper indexing
- No built-in rate limiting or request throttling on the gateway side
- Provider API keys must be exposed in environment variables for auto-detection to work
- Dashboard is a React frontend only; backend metrics are stored in ClickHouse/Redis
Perfil de infraestructura
Estimación inicial derivada del repositorio y de los servicios detectados.
Despliegue detectado
Qué señales reales de instalación hemos encontrado.
Run `docker-compose up -d` from the project root. The gateway listens on port 8080, scoring runs locally, and the dashboard starts at :3000. For production use, replace localhost with your server's IP or domain name in environment variables. Ensure ClickHouse is running for trace storage and Redis for metrics counters.
Stack y dependencias
Componentes inferidos a partir de manifests, configuración y documentación.
Puede sustituir a
Relaciones detectadas por el análisis del repositorio y su categoría.
Experiencia real de la comunidad
Ayuda a contrastar la estimación automática con instalaciones reales.
Todavía no hay datos comunitarios para este proyecto.