This thesis deals with the design and development of a production alert analysis and response system using AI, a serverless event-driven system that automates the response to infrastructure alerts generated by New Relic. In traditional incident management, the on-call team manually collects data from multiple sources (metrics, logs, service topology) before being able to make a diagnosis, with average response times on the order of tens of minutes. The system implemented replaces this collection phase with an autonomous AI agent. The architecture is entirely serverless on AWS: a New Relic alert passes via EventBridge to a Lambda that forwards it to the AgentCore runtime, which performs the investigation. The agent, based on the Strands SDK framework and implemented using AWS Bedrock AI models, accesses 36 MCP tools: 26 from New Relic for metrics and alert queries, 10 from AWS for querying cloud resources, and produces a structured report with root cause, impacted services, and remediation actions. The result is published to an SNS topic and delivered to the team via an Adaptive Card on Microsoft Teams sent by a dedicated lambda and accompanied by a Mermaid topological diagram and numbered remediation steps. This allows the MTTR to be knocked down from tens of minutes to less than 15.
Questa tesi tratta la progettazione e lo sviluppo di un sistema di analisi e risposta ad alert di produzione tramite AI, un sistema serverless event-driven che automatizza la risposta agli alert di infrastruttura generati da New Relic. Nella gestione tradizionale degli incidenti il team on-call raccoglie manualmente i dati da molteplici sorgenti (metriche, log, topologia dei servizi) prima di poter formulare una diagnosi, con tempi medi di risposta dell'ordine di decine di minuti. Il sistema realizzato sostituisce questa fase di raccolta con un agente AI autonomo. L'architettura è interamente serverless su AWS: un alert New Relic transita tramite EventBridge ad una Lambda che lo inoltra al runtime AgentCore, che esegue l'investigazione. L'agente, basato sul framework Strands SDK e implementato utilizzando modelli AI di AWS Bedrock, accede a 36 strumenti MCP: 26 di New Relic per query su metriche e alert, 10 di AWS per interrogare risorse cloud e produce un report strutturato con causa radice, servizi impattati e azioni di remediation. Il risultato viene pubblicato su un topic SNS e consegnato al team tramite un'Adaptive Card su Microsoft Teams, inviata da una lambda apposita e corredata di diagramma topologico Mermaid e passi di remediation numerati. Questo permette di abbattere l'MTTR da decine di minuti a meno di 15.
Sistema di analisi e risposta agli alert basato su intelligenza artificiale
BASSO, KEVIN
2025/2026
Abstract
This thesis deals with the design and development of a production alert analysis and response system using AI, a serverless event-driven system that automates the response to infrastructure alerts generated by New Relic. In traditional incident management, the on-call team manually collects data from multiple sources (metrics, logs, service topology) before being able to make a diagnosis, with average response times on the order of tens of minutes. The system implemented replaces this collection phase with an autonomous AI agent. The architecture is entirely serverless on AWS: a New Relic alert passes via EventBridge to a Lambda that forwards it to the AgentCore runtime, which performs the investigation. The agent, based on the Strands SDK framework and implemented using AWS Bedrock AI models, accesses 36 MCP tools: 26 from New Relic for metrics and alert queries, 10 from AWS for querying cloud resources, and produces a structured report with root cause, impacted services, and remediation actions. The result is published to an SNS topic and delivered to the team via an Adaptive Card on Microsoft Teams sent by a dedicated lambda and accompanied by a Mermaid topological diagram and numbered remediation steps. This allows the MTTR to be knocked down from tens of minutes to less than 15.| File | Dimensione | Formato | |
|---|---|---|---|
|
thesis.pdf
accesso aperto
Dimensione
2.47 MB
Formato
Adobe PDF
|
2.47 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/116474