Luis Brancher

Looking for a junior DevOps, SRE or cloud role

Luis Brancher

DevOps, SRE and Cloud

DevOps intern, focused on observability. I run a whole homelab as code, with Terraform, Ansible and GitOps, Prometheus alerting, backups with a tested restore and every failure written down with cause and fix. Before this, 15 years in audiovisual, where I coordinated a team of 13.

Now
DevOps intern since Sep 2026
Public code
7 public repositories · 15 technical articles
Stack
Terraform, Ansible, Kubernetes, ArgoCD, Linux, Prometheus, Grafana, GitHub Actions, AWS
Education
Systems Analysis and Development (in progress)
Degree in Advertising
Languages
Native Portuguese, English B2 (moving to C1)
Citizenship
Brazilian and Italian (EU)
Based in
Blumenau - SC - Brazil
01

Stack

Grouped by area, with where each thing was used and how far it goes.

AreaToolsUsed in
Infrastructure as codeTerraform, HCP Terraform, Ansible, Ansible Vault
Kubernetes and GitOpsKubernetes (k3s), ArgoCD, Helm, cert-manager, CloudNativePGA single node, for my own use.
Storage and backupTrueNAS, ZFS, NFS, pg_dump, rclone
CloudAWS (EC2, VPC, IAM), GCP (study), CloudflareFundamentals with a real project: one AWS instance for about six months, now turned off. GCP just starting.
CI/CDGitHub Actions, Forgejo Actions, GitOn GitHub, automated deploys with no test stage. On Forgejo, my own runner publishes 6 internal sites on every push, and on two of them a validation check blocks the deploy when it fails. The runner lives in homelab-infra, which is public; the site repos are private.
ObservabilityPrometheus, Alertmanager, blackbox_exporter, Grafana, Loki, ScrutinyMy own PromQL rules, outside-in probes and backup alerts on the age of the last success, all on Telegram and tested by simulating the failure. Loki receives no logs yet. At the internship, alert analysis.
Networking and securityOPNsense, VLANs, Firewall, Tailscale, WireGuard, SSH hardening, nginxIn lab work, at PUC-PR: SSH brute-force detection with Wazuh and orchestrated response in Shuffle, and Snort rules assessed for false positives.
Systems and containersLinux (Debian), Proxmox, LVM-thin, systemd, Docker ComposeDebian on VMs and LXC, disks by UUID and by-id, LVM-thin with fstrim. Docker with Compose only, no images of my own.
ScriptingBash, PythonScripts and automation, not software development.
02

Projects

What I built and what each project proves. Decisions and what is still missing open inside each one.

GitOps cluster on the homelab

A git push updates the service and reverting the commit rolls it back. k3s with ArgoCD running Nextcloud, Immich and Jellyfin, with automatic TLS and access only through the Tailscale network. The first deploy went through more than 10 failures before all five apps were Synced and Healthy.

k3sArgoCDHelmcert-managerCloudNativePGTailscale

Decisions and what is missing

Decisions

  • App of Apps: a single manual bootstrap, and every new service comes in as a file in the repo.
  • One Helm release per service: isolated failures, granular rollback and a clean diff.
  • Postgres on local disk, off NFS, because of fsync and locking. The cost showed up later: the databases were left out of the snapshots.
  • Secrets stay out of git: the repo only holds placeholders documenting the structure.

Still open

GitOps covers only part of the cluster, and a full recovery still takes about 10 manual steps. Next: SOPS, sync waves, pinned versions and staging.

Alerting and outside-in monitoring

Prometheus and Alertmanager outside the cluster, notifying on Telegram about a host down, a disk full or filling up, memory, a site down, a certificate expiring and an overdue backup. A blackbox exporter probes the services the way a user would, with an uptime dashboard in Grafana as code. While building the backup alert, I found that the scheduled Git dump had never run.

PrometheusAlertmanagerblackbox_exporterPromQLGrafanaAnsible

Decisions and what is missing

Decisions

  • Monitoring lives outside what it monitors: if the cluster goes down, the metrics of the outage are still there.
  • Backups are alerted on the age of the last success, which the script itself writes to the textfile collector, not on the error: a script that dies silently lets the value age. An absent() covers the case where it never ran.
  • Every rule tested by simulating the failure without taking the service down: the site pointed at 127.0.0.1, the timestamp aged by 27 h and an exporter stopped. In all three, the alert and the resolve arrived.
  • The blackbox exporter listens on localhost only: /probe accepts any target and, exposed, would turn the machine into a network scanner.
  • Config validated with promtool and amtool before Ansible installs it, and the disk warning inhibited when the critical for the same disk fires.

Still open

Loki receives no logs yet, the cluster and the databases have no metrics of their own and there is no SLO. Next are an external check for when monitoring itself goes down and a logs and metrics agent on k3s.

Backups with a tested restore

The databases went four months with no backup, and nothing complained. Now there is a validated daily dump, ZFS snapshots, replication and an encrypted off-site copy. Restoring into a throwaway Postgres brought back 7,183 assets and 7,124 embeddings, matching production.

CronJobpg_dumpCloudNativePGTrueNASZFSrclone

Decisions and what is missing

Decisions

  • The dump writes to .tmp, validates with pg_restore --list and only then renames: a half-written file never looks valid.
  • Fixed order: dumps first, then the snapshot that includes them, then the off-site copy.
  • COPY instead of SYNC for the off-site copy: an accidental delete or ransomware does not reach it. The trade-off is having no versions.
  • Striped pools on purpose, for more space, with the risk accepted in writing.

Still open

Restoring from the off-site copy is proven for the dumps: they came back from Drive byte for byte, with the keys kept off the NAS. A full end-to-end recovery has not been exercised yet.

Homelab infrastructure as code

Proxmox VMs and containers provisioned with Terraform and configured with Ansible, with the network (OPNsense, 4 VLANs, switch and AP) in its own repo. About 3,600 versioned lines across three repos.

TerraformHCP TerraformAnsibleProxmoxOPNsense

Decisions and what is missing

Decisions

  • HCP Terraform Agent inside the network: jobs run over an outbound connection, with no open ports.
  • Moved from telmate/proxmox to bpg/proxmox after two compatibility errors in a row with Proxmox 9.
  • Least-privilege API token and unprivileged LXC.
  • Network validated on a bench before moving the house over, and configured by hand before becoming a playbook.

Still open

Repeated tasks not yet turned into roles, unpinned k3s and ArgoCD versions, and a NIC fix that lives only on the host.

Self-hosted Git with continuous deploy

A push to main publishes the site, and a validation error blocks the deploy. Forgejo on its own LXC, created with Terraform and configured with Ansible, with an Actions runner that publishes 6 internal sites. I ran the negative test: a link broken on purpose turned the job red, the upload was skipped and the live site did not change.

ForgejoForgejo ActionsTerraformAnsibleCaddyrsync

Decisions and what is missing

Decisions

  • Deploy key locked to rrsync: no shell, and it only writes to the sites folder. The cost is no isolation between sites on the same machine.
  • Git lives outside the cluster, and the infrastructure code stays on GitHub: Git does not become a dependency for rebuilding the cluster it describes.
  • One LXC per audience, because Tailscale sharing is per machine: whoever sees the travel site cannot reach the finance one.
  • Unprivileged runner, registered offline, with the secret generated on the server. None of it goes through the repo, which is public.

Still open

There is no alert when a runner job fails, and validation only exists on two of the six sites. The daily Forgejo dump now has an age alert, but restoring it has not been tested yet. The site repos are private; the runner code is public.

First IaC: this website on AWS

For about six months this site ran on a Debian server on AWS, created with Terraform, configured with Ansible and published by GitHub Actions via rsync over Tailscale. In Sep 2026 I moved it to GitHub Pages: the site is static and did not need a server. The infrastructure code is still public.

AWSTerraformAnsibleGitHub ActionsCloudflarenginx

Decisions and what is missing

Decisions

  • Custom VPC and IAM with an explicit Deny for instances other than t4g.micro, because a Deny beats any policy attached later.
  • A key pair just for CI, separate from my personal key, after four failures in a row on the first workflow.
  • Turning the infrastructure off once it stopped paying for itself: a static site does not justify a running server.

Still open

It was a single instance, with HTTPS terminated at Cloudflare and the origin on HTTP, and the workflow only deployed, with no tests. Today I would close the origin with TLS and swap the long-lived key for OIDC.

03

What has broken

Real failures, my own included, with cause and fix. Where a public record exists, the row links to it.

SymptomCauseFixPost
Daily Git backup never produced a filecron runs with a minimal PATH, and the script called a binary outside it. It died with exit 127 every day since the first night, and the error went to an email nobody receives. It showed up when I set out to write the alert for that backup.The script now sets its own PATH and records the time of the last success. If it gets older than 26 hours, or never shows up, an alert reaches Telegram.Read
Nightly backup queues 57 GiB already at the destinationA 403 quota error made one folder listing fail, and rclone treated "could not read" as "does not exist". Three of my hypotheses fell before the log answered.Run aborted, my own OAuth credential and an exclude list. The scheduled run went from over 9 hours to about 8 minutes.Read →
Proxmox thin pool at 98.7%, with no alertVM disks with discard=ignore: the VM knew what it had deleted, but Proxmox did not. With media being downloaded and deleted, the pool only grew.discard=on, fstrim and a timer. The system volume dropped to 56.4%, with 48 GB recovered.Read →
Guest Wi-Fi reaches the router adminThe block covered the other VLANs, but not the IPs of the router itself. The flaw was mine and dated back to May.Rules on (self), allowing only DNS, DHCP and NTP, then a real test from a phone on the guest network.Read →
Server drops off Tailscale, 10,000 km from homeAfter a power outage. Hypothesis: a static IP and DHCP conflict, plus the ISP router. The cause was never proven.Got in through the LAN from the workstation, then SSH over IPv6 link-local. It became the reason to rebuild the homelab.Read →
Proxmox host freezes under heavy trafficAn e1000e/I219 NIC bug. I first suspected the CPU and went up to 16 vCPUs, with no effect.Offloads turned off, moving the work to the CPU. It lives outside the code, so it is recorded as drift.Read →
Scrutiny flags the boot SSD as Failed121 old media errors, from an SSD bought used. smartctl showed 100% spare and clean tests.From "replace now" to "watch it": the counter never resets, and what matters is whether it grows. Config kept off the NAS.Read →
metadata.annotations: Too long: may not be more than 262144 bytesCNPG CRDs too large for the last-applied-configuration annotation.ServerSideApply=true in the Application syncOptions.Read →
rsync: connection unexpectedly closedFirst, the key in the secret was missing its final newline. Then the passphrase, which the runner cannot type.A key pair dedicated to CI, separate from my personal key.Read →
04

How I work

Four rules I follow on the homelab and bring to work.

  1. Everything goes through git

    On the cluster, a push updates the service and a revert rolls it back, with ArgoCD applying what is in the repo. Network and machines are in code too. Anything sensitive stays off GitHub, in a Forgejo instance I host on a homelab LXC, with its own Actions runner.

  2. AI as a pair, with the decision on me

    When a tool is new to me, I do it on my own to understand what is going on. Then I use AI as a teacher, to answer questions, and to speed up what I already know how to do, always reading the diff before applying it.

  3. I only trust what I have tested

    A backup only counts after a restore, and a playbook runs twice to prove the second run changes nothing. Security works the same way: VLANs, least-privilege IAM and access through Tailscale, and the audit still found guests reaching the router admin.

  4. I write down the failure along with the fix

    I publish technical articles from my own notes, with help from AI on the writing, and keep runbooks in the repos. It is what I already did in audiovisual, coordinating creatives, production companies and clients.

05

Homelab

Real hardware, a segmented network and services in daily use, described in code. It is a single node for my own use, with no external traffic: not high availability.

From bare metal to service

  1. TerraformCreates VMs and LXC on Proxmoxhomelab-infra
  2. AnsibleConfigures the OS and brings up k3shomelab-infra
  3. ArgoCDReconciles the cluster with githomelab-gitops
  4. HelmShips one release per servicehomelab-gitops
  5. PrometheusMetrics, probes and Telegram alerts, outside the cluster on purposehomelab-infra

Segmented network

VLANIDAccess
trusted10Unrestricted
iot20Internet only
guest30Internet only
lab40Internet and trusted

Machines

MachineRole
Mini PC N150OPNsense: router, firewall and VLANs
Lenovo M80q, i5-13500TProxmox: k3s VM, monitoring and Forgejo LXCs
Lenovo V50s, i3-10100TrueNAS: 2 data SSDs and 2 backup HDDs
Raspberry Pi 4Omada Controller, MotionEye and UpSnap

Services

ServiceRoleHost
nextcloudFile Synck3s
immichPhoto Backupk3s
jellyfinMedia Serverk3s
cert-managerTLS / Let's Encryptk3s
cnpg-operatorPostgreSQL Operatork3s
prometheusMetricsLXC
alertmanagerTelegram AlertsLXC
blackboxHTTP and DNS ProbesLXC
grafanaDashboardsLXC
lokiLog AggregationLXC
forgejoSelf-hosted GitLXC
forgejo-runnerCI/CD RunnerLXC
scrutinyDisk HealthTrueNAS
omada-controllerNetwork MgmtPi4
motioneyeNVR / CCTVPi4
upsnapWake-on-LANPi4

Inventory declared in code, not live monitoring.

View the full network topology
trusted · VLAN 10iot · VLAN 20guest · VLAN 30lab · VLAN 40WiredWi-FiWANVLAN IoT (ID 20)VLAN Guest(ID 30)VLAN Lab (ID 40)VLAN Trusted (ID 10)Raspberry Pi 4TrueNASLenovo V50s · i3-10100Proxmox Host (i5-13500T)VM k3s-node · k3s + ArgoCDLXC lxc-monitoringLXC forgejoWireGuard TunnelPoE+NFSInter-VLAN AccessInternetISP Modem (Bridge Mode)Tailscale VPNOPNsense FirewallMiniPC N150 · 4 NICs2.5GbE SwitchOmada WiFi 6 APTVsSmart devicesPrinterPi 3Cudy TR3000OpenWrt · APWorkstationPCSmartphonesLaptopsScrutinyOmada ControllerMotionEyeUpSnapnextcloudimmichjellyfincert-managercnpg-operatorprometheusalertmanagergrafanalokiblackboxforgejoforgejo-runner
06

Career

I started with a homelab out of curiosity and studied cybersecurity to understand what I was doing. That is where I fell in love with infrastructure. Today I treat everything as code and think about protection before spinning up any resource. I come from audiovisual, where documentation, troubleshooting and tight deadlines were already part of the job.

Experience

  1. Sep 2026 – now

    DevOps Intern

    ZPE Systems

    Focused on observability.

    • Alert analysis: I read the firing data and propose what to tune or remove, describing the risk of each change.
    • On my own initiative, I built an interactive map of the flows of an internal platform, now used by the team as documentation.
  2. Mar 2026 – now

    Personal cloud and DevOps projects

    Public code, on a homelab foundation since 2023

    • Homelab rebuilt as code, from bare metal to service, in public repositories.
    • PostgreSQL backups with verification and a tested restore.
    • Network split into 4 VLANs with OPNsense, configured through Ansible and the API.
    • Self-hosted Git with an Actions runner that validates and publishes sites on every push.
    • First IaC on AWS: Terraform, least-privilege IAM and deploys through GitHub Actions over Tailscale.

Before: audiovisual

  1. 2021 – 2026

    Freelance motion designer and editor, for clients in Brazil and abroad

    Alicerce Educação and international clients

    Over 4 years at Alicerce, creating the videos and hiring local production companies across Brazil, with 70 recorded projects. In 2026, motion design, AI-assisted spots and podcast editing for clients abroad.

  2. 2019 – 2021

    Creative Lead

    Henry Schein Brazil

    Promoted from video editor to coordinator in under a year. A team of 13, 3 videographers and 10 designers, running monthly retail, e-commerce and video campaigns.

  3. 2016 – 2019

    Marketing Analyst, then freelancer

    Dental Cremer (now Henry Schein) and agencies

    First video professional hired: I created the video department and ran it alone for about a year and a half.

  4. 2010 – 2016

    TV and production: editing, live image direction and motion

    RBS TV, RBA TV, UNIASSELVI and Fábrica de Cinema do Alto Vale

Certifications and courses

NameIssuerDate
LFEL1009: Getting Started with OpenTofuLinux FoundationNov 2026
LFS158: Introduction to KubernetesLinux FoundationSep 2026
Terraform Basics Training CourseKodeKloudApr 2026
Fortinet Certified Associate in CybersecurityFortinetAug 2025
Fortinet FortiGate 7.6 OperatorFortinetAug 2025
Cyber Security 101 CertificateTryHackMeJun 2025
Cybersecurity Professional CertificateGoogleMay 2025

Studying for, no exam taken yet: Terraform Associate (004) and Google Associate Cloud Engineer. All credentials, verifiable, on Credly

Education

  1. Jul 2026 – Jul 2028, in progress

    Systems Analysis and Development

    UniCesumar

  2. May – Dec 2025

    Completed Cybersecurity courses

    PUC-PR

    Linux, Python, CCNA CyberOps, Network Defense and Ethical Hacking. Labs on Wazuh with Shuffle, Snort and pfSense.

  3. 2009 – 2013

    Advertising; Graphic Design technologist degree

    UNIASSELVI

  4. 2023

    General English, 8 weeks in Liverpool

    Kaplan International

07

Blog

Homelab, AWS and automation, written while the problems were happening.

All articles →

Hiring a junior for DevOps, SRE or cloud?

I reply by email or LinkedIn.

luisbrancher@proton.me