StackWatchProduction errors, localized at the root.

AI-driven root cause analysis for Java production errors — stacktrace fingerprint merging + LLM root cause localization + scheduled weekly digest. Cut mean time to localize incidents by ~40%.

Quick Start

GitHub

Three-tier cascade

L1 fingerprint cache → L2 vector merge → L3 LLM root cause. Only ~1% of exceptions actually call the LLM — the cost ladder cuts analysis cost by two orders of magnitude.

Bidirectional feedback flywheel

Developer confirmations flow back as few-shot positive samples and anti-pattern negative samples, both injected into the next L3 prompt. Every call gets smarter.

Zero-infrastructure start

Starts with no DB, no Kafka, no vector store. L1/L3 work out of the box. L2 (PgVector) and Kafka unlock progressively via config flags.

Architecture

Five layers. One pipeline.

01 / Collect

Collector

Logback Appender · HTTP /collect · Kafka (off by default)

02 / Preprocess

Fingerprinter

SHA-256 + framework-frame filtering · versioned fingerprints

03 / Analyze

Analyzer — L1 / L2 / L3 cascade

Exact cache hit → vector merge → LLM root cause (core hub)

04 / Aggregate

Aggregator

Real-time surge detection · weekly Top-N aggregation

05 / Notify

Notifier

Feishu real-time alert · Feishu weekly report

Start where your errors are.

Read the docs →