Streaming Observability Platform with Local LLMs

Problem

Observability pipelines generate high-volume event streams, but turning raw metrics into actionable incident context still requires human triage. This project explores whether local LLM inference can close that gap — without sending telemetry to external APIs.

Architecture

• Ingest: Kafka topics carrying synthetic + real event streams
• Process: Spark Structured Streaming — tumbling windows, watermarks for late data, foreachBatch materialization into DuckDB
• Detect: rule-based anomaly detector with tiered severity levels
• Summarize: local LLM (Ollama) generates incident summaries from anomaly context

Results

Full write-up: From Hello World to Streaming Insights on Medium.

Detailed case study with diagrams coming soon.