I build infrastructure and orchestration for AI workloads, with 20+ years of experience building, delivering and operating software at scale. Currently Principal Software Engineer at Midokura, in Barcelona. If you work on AI infrastructure or software delivery, I would like to hear from you: anglorvaroa@gmail.com

Projects

• GPU infrastructure for AI workloads Midokura, now

  Systems to provision, manage and operate GPU infrastructure and high-bandwidth fabric in private datacenters. Multi-tenant IaaS and PaaS layers for AI workload orchestration.

• Sony Edge AI Sensing Platform Midokura, about 3 years

  Re-architected the real-time inference streaming backend: 100x performance with a lower resource footprint. Moved release from manual QA to automated pipelines: cycle time cut by more than 50%, across EU and Japan teams.

• Entity Platform at New Relic 2020 to 2022

  The inventory of everything New Relic observes: hosts, containers, services, their metadata and relationships. Most of the observability product depended on it. It was on the critical path of 50 to 75% of company OKRs, under a 99.8% SLA. Led the design of the telemetry analysis system that processed millions of datapoints per second and inferred entity types and relationships from open rules.

• Internal developer platform for 1,500 engineers at Adevinta 2016 to 2020

  Kubernetes PaaS for development, delivery and operations, used by marketplaces across the group in Europe and South America. Led the migration of existing workloads from on-premise and EC2 onto it. Developer tooling from zero to hundreds of users.

• Network virtualization for OpenStack Midokura, 2013 to 2016

  SDN control and data plane: L2 and VxLAN gateways, NAT, hardware VTEP integration.

• Music catalogue ingestion at Last.fm 2007 to 2009

  Managed the systems that ingested and served the largest streaming music catalogue in the world at the time, with 60M tracks, 40M users.

Full history on About.

Writing

Software delivery at AI volume

• DDoSing software delivery pipelines2026
• Inferencing delivery bottlenecks away2026
• Code reviews can't keep up2026
• AI-generated code will choke delivery pipelines2025

Infrastructure for AI

• AI workloads challenge the cattle model2026

Platforms and operations

• PoC is a framework of perverse incentives2026
• Why aren't we all serverless yet?2025
• Identifiers are better off without meaning2024
• Alert on symptoms, not causes2024
• How about we forget the concept of test types?2024
• How organisations cripple engineering teams with good intentions2024
• Migrating an Eureka-based microservice fleet to Kubernetes2020
• How to build a PaaS for 1500 engineers2020
• Kubernetes made my latency 10x higher2019

JVM internals

• Sizing Kubernetes pods for JVM apps without fearing the OOM Killer2019
• GC forensics by example: multi-second pauses and allocation pressure2018
• How does the default hashCode() work?2017
• DirectBuffer creation and disposal has hidden contention on sun.misc.Cleaner2016

Talks

• How we built a PaaS for 1,500 engineers
  The Adevinta platform: why we built it, how, and how we measured it.PlatformCon, 2022
• Salesman tricks for the Platform Engineer
  How the Adevinta platform team got other teams to adopt the platform.PlatformCon, 2022
• Frugal memory management on the JVMMeetup, 2016