De los dashboards a las decisiones: integración de IA en operaciones DevOps a gran escala
Los equipos DevOps modernos no carecen de datos; carecen de contexto operacional unificado.
Esta charla comparte nuestra experiencia introduciendo AI-assisted incident investigations en un entorno global de e-commerce. Exploraremos cómo la IA puede correlacionar información de Kubernetes, AWS, Argo CD, plataformas de observabilidad, repositorios e historial de deployments para ayudar a tomar decisiones más rápidas y mejor fundamentadas.
También abordaremos los safeguards esenciales: read-only access, validation, auditability, stopping conditions y human oversight, incluyendo qué funcionó, qué falló y qué tareas operacionales deberían seguir siendo determinísticas.
Modern DevOps teams rarely suffer from a lack of data. The real problem is fragmentation.
During an incident, engineers may need to move between Kubernetes, AWS, Argo CD, observability platforms, CI/CD pipelines, repositories, edge services, and internal documentation. Each system provides part of the answer, but reconstructing the full operational story still requires time, experience, and repeated manual investigation.
This talk presents our experience introducing AI-assisted operational workflows into a global e-commerce environment running distributed services across multiple platforms and environments.
Rather than using AI only to generate commands or summarize logs, we explored how it could act as an investigation layer across existing DevOps systems: collecting context, correlating events, proposing hypotheses, and helping engineers move from raw telemetry to an informed operational decision.
What the session will cover
- The operational challenges of a large-scale e-commerce platform with multiple services, environments, deployment pipelines, and observability tools.
- How operational context can be gathered from Kubernetes, AWS, Argo CD, monitoring platforms, repositories, and deployment history.
- How an AI-assisted investigation loop can plan actions, call read-only tools, evaluate observations, and refine its hypothesis.
- Why deterministic workflows and AI agents solve different categories of operational problems.
- How we introduced access controls, environment boundaries, auditability, stopping conditions, and human approval points.
- What failed during implementation, including incomplete context, misleading correlations, excessive tool calls, and confident but incorrect conclusions.
- How tracing and evaluation helped us understand whether the system was actually improving incident investigation.
- Which tasks proved useful for AI assistance and which remained safer and more reliable as traditional automation.
A central theme of the talk is that connecting a language model to operational tools is the easy part. The difficult part is designing a system that knows:
- What it is allowed to inspect.
- How to validate what it finds.
- When to stop.
- When to defer to an engineer.
The session will include a practical architecture, an example incident-investigation flow, the controls required for enterprise use, and the lessons we learned while applying these ideas in a real production environment.
Key takeaways
Attendees will leave with:
- A realistic framework for deciding where AI can add value in DevOps.
- A clearer understanding of where deterministic automation remains the better choice.
- Practical guidance on the technical safeguards required before introducing agents into operational workflows.
- Lessons learned from applying AI-assisted investigation in a real enterprise environment.
I am a Cloud Architect and DevSecOps specialist with more than 20 years of experience in software development, cloud architecture, platform engineering, automation, and large-scale digital operations.
Throughout my career, I have worked in complex international environments supporting global companies and business-critical platforms, helping teams design, modernise, and operate distributed systems across cloud infrastructure, Kubernetes, CI/CD, GitOps, observability, security, and reliability engineering.
A major part of my current work focuses on the practical integration of artificial intelligence into software delivery and operations. I design and evaluate AI-assisted workflows, agentic systems, and operational agents capable of gathering context from multiple platforms, using tools, analysing intermediate results, refining hypotheses, and supporting engineers during incident investigation and technical decision-making.
My interest in agentic AI goes beyond conversational assistants. I focus on the architecture required to make these systems useful in enterprise environments: tool orchestration, context management, access boundaries, validation, auditability, stopping conditions, human oversight, and the balance between autonomous reasoning and deterministic automation.
I am also a qualified teacher with more than 15 years of training experience. I have delivered technical education for companies and professional teams, as well as meetups, webinars, workshops, and sessions at venues including Google Campus Madrid.
My goal as a speaker is to turn complex technical subjects into clear, practical lessons grounded in real production experience, including not only what worked, but also the limitations, failed approaches, and engineering decisions behind the final result.