AI-powered API observability: a Kong + LLM case study
We piped Kong gateway logs through a model-backed anomaly summariser to cut the noise in our observability stack. Here are the real numbers and the prompt that worked.
Signal over volume
Raw gateway logs are high volume and low signal. Summarising them with a model turned thousands of lines into a short, ranked list of what actually changed, with a meaningful drop in time-to-detect for the incidents we replayed.
The trick was constraining the prompt to known failure shapes and feeding it structured fields rather than raw text. Let the model rank and explain; do not let it invent categories.
More from the FlorenceNext team
- Tech10 Mar 2026FlorenceNext Team
Master Data Management without the multi-year programme
Why most MDM efforts fail, and a smaller-scoped approach that gets you to a single source of truth in a quarter.
Tech24 Feb 2026FlorenceNext TeamUsing Claude as a pair-programmer on a real MuleSoft project
Notes from three months of treating an LLM as a junior on the team. What works, what to keep humans on, and where the savings actually land.
Tech10 Feb 2026FlorenceNext TeamCost-aware MuleSoft on AWS: the bits we wish someone had told us
Patterns that bring runtime fabric cost down without sacrificing the latency budget your SLAs depend on.