DeepWatch Sentinel is an advanced AI Agent system that acts as the "brain" for infrastructure observability. It leverages local Large Language Models (LLMs) to perform autonomous reasoning over statistical data, providing human-like verdicts on system health and experiment results.
- Agentic Reasoning: Implements ReAct logic (Reasoning + Acting) to analyze complex SQL datasets.
- Dynamic Context Injection: Fuses real-time metrics from
StatGuard-Metricwith LLM prompts for grounded, data-driven insights. - Local AI Privacy: Fully powered by Ollama (Llama 3), ensuring that sensitive data never leaves the local environment.
- Actionable Verdicts: Provides business-ready recommendations (e.g., "Deploy V2 update based on +2.83% lift").
- AI Core: Ollama (Llama 3 8B)
- Persistence: SQLite (StatGuard Integration)
- Orchestration: Python 3.11, Requests
- Architecture: RAG-style Context Augmentation
- Ensure Ollama is running:
ollama serve - Setup DB: Ensure
data/experiments.dbfrom StatGuard-Metric is available in the root. - Run Sentinel:
python main.py
Sentinel uses a strict system prompt engineering strategy to ensure that the AI only interprets data present in the database, preventing "hallucinations" and ensuring 100% reliable technical verdicts.
- Multi-Agent Collaboration: Integrating specialized agents for cost optimization and security.
- Slack/Discord Integration: Automated incident reporting via AI-generated summaries.
- Advanced RAG: Vector database integration for searching through technical documentation.
Feel free to fork this project, report issues, or suggest new Agentic patterns.