Show HN: ML condenses billions of logs into a tiny snapshot your LLM can debug
TL;DR
Rocketgraph is a self-hosted open-source project for log clustering and streaming anomaly detection. It sits next to existing observability stacks such as Datadog, Loki, CloudWatch, Sentry, ClickHouse or New Relic. According to the README, the ML engine itself is LLM-free: Drain3 extracts structural log templates, Isolation Forest scores unusual templates per service, and Half-Space-Trees score fresh logs in real time.
Nauti's Take
The HN pitch sells this as LLM debugging, but the useful move happens one step earlier: logs have to be compressed into something a person or model can reason about. Rocketgraph earns points for keeping the core deterministic and using LLMs only as an explanatory layer.
The proof is still thin: a 2M-log demo does not replace comparisons against existing anomaly detection tools, false-positive rates, and real postmortems. Still, the direction makes sense because grep-style incident work breaks down faster when AI is shipping more code.
Briefingshow
Observability is becoming a bottleneck for AI-assisted software teams: code changes faster, logs get noisier, and classic dashboards often show only what someone already knew to ask for. Rocketgraph targets a real pain point by condensing raw logs before an engineer or LLM starts debugging. The decisive test is not the demo number, but whether its clusters are precise enough during real incidents.
Sources