Perfinsights

PerfInsights starts with evidence of where running software spends computing time. Uber Engineering uses production profiles to select each service's most expensive functions in Go, then asks a language model to look for known inefficient coding patterns in their source code. This focuses analysis on code that actually consumes computing resources.

Before suggestions go onward, a jury of additional language models independently checks whether each flagged pattern is present, and rule-based validators called LLMCheck look for specific mistakes, such as reporting a repeated computation inside a loop when the variable is actually defined outside it. Uber reports reducing false positives from over 80% to the low teens. The account describes how measured resource use and targeted validation narrow the findings that developers see.