Gustavo Nobrega DevOps & Infrastructure

Castlegar, British Columbia — Canada

Gustavo Nobrega

DevOps and cloud infrastructure engineer. I have spent about two decades on the unglamorous half of computing — the servers, clusters, databases and radio networks that other people’s software runs on top of.

Today I run infrastructure at infinitii ai, where machine learning meets municipal water and wastewater systems — including the self-hosted LLM gateway and agent tooling the AI work runs on. Before that: a private cloud behind a nine-million-user social network in Brazil, database clusters for banks, and a physics degree that started with a Beowulf cluster built out of spare desktops in 1997.

Portrait of Gustavo Nobrega
Role
DevOps & Cloud Engineer
Callsign
VA7PQP
Languages
PT · EN · ES · FR
1997First Beowulf cluster at UNESP Bauru
9M+Users on infrastructure I designed and ran
28 yrsVolunteer sysadmin for a research HPC lab

Hybrid cloud, treated as a cost strategy

Going all-in on public cloud is the expensive default for steady-state workloads. I put predictable baseline load on dedicated hardware — where cost per unit of compute and storage is a fraction of on-demand pricing — and reserve the public cloud for what it is genuinely good at: elasticity, managed services and burst.

Getting that split right cut data-centre cost by 60% while doubling capacity, with no loss of reliability or performance. The result was published as a case study by iWeb Technologies. It is the same discipline I run today at infinitii ai, across Microsoft Azure and dedicated servers.

01 What I’m doing now

Smart water infrastructure

At infinitii ai (formerly Carl Data Solutions) I own the platform side: a hybrid estate spanning Microsoft Azure and dedicated hardware, tuned so that reliability and cost pull in the same direction rather than against each other.

Sensors and LoRaWAN

I write firmware for environmental data loggers and climate stations, and I build the LoRaWAN networks that carry their payloads — gateways, ChirpStack, and the bridges that land the data in a backend somebody can actually query.

Search and rescue

President of Castlegar Search and Rescue. The engineering habit follows me into the field: most of my recent side projects are tools that make ground search teams faster.

02 AI infrastructure & LLM orchestration

The plumbing under the AI work: one gateway in front of every model, agents that keep working when a provider does not, and credentials that never leave the vault.

Self-hosted · one API

LLM gateway

Designed and operate a gateway that puts several cloud AI providers and local open-source models behind a single API. Load balancing across backends, automatic failover when one starts refusing work, and rate-limit management so a burst of agent traffic gets absorbed instead of dropped.

opencode · Oh My Pi

Coding-agent harnesses

Integrated multiple AI coding-agent harnesses with that gateway and tuned the routing: which model takes which role, weighed against cost, capability and how much context the job actually needs.

Root-caused from logs

An upstream bug, fixed without the vendor

Third-party proxy software in the request path was silently corrupting request handling under load — nothing raised an error, the behaviour was just wrong in a way that looked like a model problem. I traced it through log analysis to a bug in the proxy itself and fixed it with no vendor support.

Customer-facing, in product

“Carl” bot platform

A self-hosted conversational AI that answers customers inside a commercial web product. I extended the open-source messaging-bot framework underneath it — voice and audio output among other capabilities — and connected it to enterprise channels including Microsoft Teams and WhatsApp.

Isolated deployments

Credential vault

A vault that brokers API keys and secrets out to several isolated agent deployments, so internal and external-facing environments share tooling without ever sharing credentials.

Open source · MCP

MCP server for SAR operations

Authored and published mcp-d4h, a Model Context Protocol server that connects Claude Desktop to the team-management platform search-and-rescue organisations run on — dozens of tools that let an assistant query and act on live operational data.

03 Experience

Twenty-plus years across Canada and Brazil.

  1. Jun 2018 — Present

    Director of IT · DevOps & Cloud Engineer

    infinitii ai Inc. / formerly Carl Data Solutions / British Columbia, Canada

    • Own every project and system with reliability and performance as the standing brief.
    • Built and run a hybrid cloud across Microsoft Azure and dedicated servers, cutting spend without trading away reliability.
    • Designed and operate a self-hosted LLM gateway: several cloud AI providers and local open-source models behind one API, with load balancing, failover and rate-limit management for heavy, bursty agent traffic.
    • Root-caused a third-party proxy bug that was silently corrupting request handling under load — found it through log analysis and fixed it upstream without vendor support.
    • Built and run “Carl”, a self-hosted conversational AI that serves customers inside the company’s commercial web product, and extended the open-source bot framework under it with voice output and Microsoft Teams and WhatsApp channels.
    • Firmware engineer for IIoT: climate stations and environmental data loggers.
    • Designed and maintain a long-range, low-power LoRa network serving the company’s IIoT fleet.
    • Backend engineering — APIs and bridge services that read from one system, reshape the data, and publish it into another.

  2. Jul 2018 — Sep 2019

    CTO · Software Engineer

    i4C Innovation Center / Trail, BC, Canada

    • Set the technology vision and kept engineering spend aligned with what the business actually needed.
    • Firmware engineer for IIoT environmental data loggers.

  3. May 2006 — Jan 2018

    Shareholder · CTO · Senior SysAdmin & DBA

    eSapiens Tecnologia SA / Florianópolis, Santa Catarina, Brazil

    • Built and operated a private cloud — database clusters, web service clusters, CDN, firewalls — behind a Brazilian social network with more than 9 million users and 150,000 concurrent at peak.
    • Optimised relentlessly for reliability, security and low cost, because all three were constraints at once.
    • Introduced the data-modelling practice: algorithms, predictive models and ad-hoc analysis over the customer database to turn large data sets into decisions.

  4. Apr 2017 — Jan 2018

    Co-founder & Technical Advisor

    Blocko Ventures / Florianópolis, Santa Catarina, Brazil

    • Developed a private blockchain solution used across the company’s projects.

  5. Jan 2011 — Dec 2014

    Co-founder · CEO · IT Consultant

    Nobrega Tecnologia da Informação Ltda / Bauru, São Paulo, Brazil

    • Consulting across Linux and Microsoft estates, with a specialty in database clusters for large Brazilian companies and banks.

  6. Nov 2007 — Jul 2011

    Co-founder & IT Consultant

    Tricore Tecnologia da Informação e Comunicação Ltda / Bauru, São Paulo, Brazil

    • Delivered Linux and Microsoft infrastructure projects, again with database clustering at the centre.

  7. Aug 2004 — Sep 2007

    SysAdmin & DBA

    Planae Tecnologia da Informação Ltda / Bauru, São Paulo, Brazil

    • Ran Linux and Windows servers for the largest loan-recovery company in South America, including a large PostgreSQL cluster.
    • Consulted on databases, mail, web and DNS infrastructure for a long list of clients.

04 Selected projects

Things I built end to end, from the enclosure to the API.

2019 — Present

Fresh Water Data Commons

A Canadian initiative digitising strategic water bodies to give a real-time view of ecosystem health. I was firmware engineer for the sensors and LPWAN engineer for the network — deploying the LoRaWAN estate and the software stack that integrates it with backend systems.

FirmwareLoRaWANIntegration

2018 — Present

Astra AirSafe

An ultra-low-cost industrial smoke sensor, developed from hardware through firmware. 56 grams, a reading every three seconds, temperature and particulates. Enclosure designed and 3D-printed in-house; firmware and API in Lua; backend on AWS.

HardwareLuaIIoTAWS

2013

Private cloud migration

Moved an entire physical fleet onto a private cloud in four stages: iSCSI storage cluster, XenServer cluster, virtual machine build-out, then service cut-over. This is the migration that produced the cost and capacity result above.

XenServeriSCSIMigration

1997

First Beowulf cluster at UNESP Bauru

Built to run chemistry simulations: Linux nodes, NFS and NIS+ for shared files and accounts, a batch queue manager for scheduling, and a MOSIX single-system-image lab alongside it. The machines are long gone; the habit stuck.

HPCLinuxMOSIX

05 Open source

78 public repositories on GitHub. A few worth pointing at.

06 Stack

Tools I have run in production, not just read about. These four are what I operate day to day.

6-node pool

XCP-ng

Hypervisor administration end to end — host lifecycle, pool membership, storage repositories and live migration.

3-node cluster · Squid

Ceph

OSD rebalancing, BlueStore fragmentation tuning and MDS cap tuning, with CephFS exported over NFS-Ganesha.

12-node cluster

Cassandra

Upgraded 3.11 → 4.1 with 5.0 next. Repair orchestrated through Reaper; backups via Medusa into MinIO.

Multi-namespace production

Kubernetes · RKE2

Rolling upgrades, node drain and PodDisruptionBudget management, and CVE patching across test and production clusters.

LLM gateway & agents

  • LiteLLM
  • Ollama
  • Local open-source models
  • Model Context Protocol
  • opencode
  • Oh My Pi
  • Claude Desktop
  • Microsoft Teams bots
  • WhatsApp

Virtualisation & storage

  • XCP-ng
  • XenServer
  • VMware
  • Ceph
  • BlueStore
  • CephFS
  • NFS-Ganesha
  • MinIO
  • iSCSI
  • Beowulf clusters
  • MOSIX

Orchestration & CI/CD

  • Kubernetes (RKE2)
  • Helm
  • Docker
  • Ansible
  • Ansible Vault
  • Terraform
  • Jenkins
  • Azure DevOps
  • Codemagic

Data & telemetry

  • Cassandra
  • Reaper
  • Medusa
  • PostgreSQL
  • MySQL
  • MS SQL Server
  • MongoDB
  • Redis
  • InfluxDB
  • Telegraf
  • Grafana
  • SQLite
  • Hadoop

Cloud

  • Microsoft Azure
  • Cloudflare Workers
  • Cloudflare Pages
  • Cloudflare R2
  • AWS
  • Apache CloudStack

Systems & networking

  • Linux
  • FreeBSD
  • NGINX
  • Apache
  • pfSense
  • Unbound
  • BIND
  • OpenVPN DCO
  • Azure Point-to-Site
  • iptables / pf
  • Postfix
  • Nagios
  • Supervisord

Windows, AD & M365

  • Windows Server
  • Active Directory
  • Exchange Online
  • DMARC & anti-phishing
  • Samba / CIFS
  • SMB3 kernel tracing

APIs & web

  • FastAPI
  • Fastify
  • Preact
  • Vite

Mobile & apps

  • Flutter
  • Riverpod
  • Drift
  • PMTiles
  • Android SDK
  • DJI Mobile SDK v5
  • PWA / OPFS
  • Maestro

Embedded & radio

  • LoRa / LoRaWAN
  • ChirpStack
  • Zephyr
  • ESP32
  • ESP8266 / NodeMCU
  • STM32
  • Arduino

Languages

  • Python
  • Dart
  • Go
  • C / C++
  • Kotlin
  • Lua
  • Bash
  • JavaScript
  • TypeScript
  • PHP
  • R

07 Volunteering

Unpaid work I would not give up.

6+ years, ongoing

President

Castlegar Search and Rescue

I chair the team as President, lead crews in the field as a Team Leader, and run incidents as a SAR Manager — the role that plans a search and decides where resources go. It began with 100 hours of ground search and rescue training at the Justice Institute of BC, and now covers advanced medical care in the backcountry, where the nearest hospital is hours away.

Team LeaderSAR ManagerAdv. Non-Urban Medical Care

Mar 1998 — Present

SysAdmin

Computer Simulation Laboratory, UNESP

Twenty-eight years of setting up and maintaining the cluster computers and software that researchers depend on — and architecting the next HPC solution so they spend their time on science instead of on queueing systems.

2024 — Present

Soccer Coach

Kootenay South Youth Soccer Association

Coaching youth soccer in the Kootenays — weekly practices through the season and games on the weekend.

Respect in Sport certified

Dec 2013 — May 2016

Code Review

OpenStack

Reviewed patches and chased down bugs in the OpenStack cloud software.

Jun 2020 — Nov 2021

Director of Web Operations

Kootenay Mountaineering Club

Kept the club’s site running so members could find hiking routes in the Kootenays and book the backcountry huts.

08 Writing

09 Education & certifications

Education

  • 2011 – 2012 Master in Materials Science POSMAT, Faculty of Science — São Paulo State University (UNESP)
  • 2007 – 2010 B.Sc. in Physics Faculty of Science — São Paulo State University (UNESP)

Languages

  • NativePortuguese
  • FullEnglish
  • FullSpanish
  • BasicFrench

Certifications

  • 2023 Amateur Radio Operator Certification — VA7PQP Innovation, Science and Economic Development Canada
  • 2022 Certified Kubernetes Administrator (CKA) The Linux Foundation
  • 2016 Linux Foundation Certified System Administrator The Linux Foundation
  • 2015 Big Data with Hadoop and NoSQL FIAP — São Paulo
  • 2015 Cloud Computing Specialization University of Illinois at Urbana-Champaign
  • 2007 Microsoft Certified Database Administrator Microsoft
  • 2006 Microsoft Certified Professional Microsoft

Get in touch

Happy to talk about infrastructure that has to stay up, sensor networks in awkward places, or anything involving mountains and radios.