English · Português · LinkedIn
Cloud operations engineer at an AWS Partner MSP: about 20 customer AWS accounts, FinOps, Terraform and eight years of on-call. I'm moving into ML platform work on Kubernetes, with the same focus on reliability and cost.
21 labs built around one notes application with an API and PostgreSQL: containers, CI/CD, Kubernetes, GitOps, observability, backup, disaster recovery, failover and chaos engineering. Each lab has code, setup and cleanup steps, and JSON test reports. Failed runs and open questions stay in the record.
Architecture · Study order · Results and evidence
| Lab | What you'll find |
|---|---|
| Chaos engineering | Pod and network faults investigated through packet captures. The documented CI run has 146 checks and five reproduced TCP connect timeouts; one backend remains unidentified. Investigation. |
| Failover and failback | A controlled switch between clusters with synchronous PostgreSQL replication, fencing of the previous writer and data checks in both directions. Validation. |
| GitOps delivery platform | From pull request to Argo CD: digest-pinned image, Cosign signature, SBOM, kind integration in CI and a Terraform EKS reference validated without apply. Evidence. |
| Disaster recovery | Velero restores the application and PostgreSQL into another cluster while the source is stopped; data is checked and both restore warnings are kept. Validation. |
| Image security | Trivy gates that must pass for adopted images and must keep blocking the vulnerable images retained for recovery. API migration. |
| Tracing | Requests traced through OpenTelemetry and Tempo, with HTTP/SQL spans and log lookup by trace ID. Validation. |
The environments run on Docker/WSL and local kind clusters that share one host, so the recovery results don't cover machine loss. I used an AI coding assistant to write much of the code and docs; the scenarios, pass criteria, validation and failure investigations are mine. The portfolio explains this and each lab's limits.
- AWS cost forecasting with a Transformer: an encoder-only Transformer written from scratch in PyTorch, served with FastAPI, trained on synthetic AWS cost series.
- RAG on AWS: Bedrock embeddings and generation, OpenSearch kNN, Lambda and Terraform.
- Model serving: an Isolation Forest behind FastAPI, with Docker, Kubernetes manifests and CI.
- Metrics over MCP: an MCP server that lets an agent run allowlisted PromQL queries against Mimir and read Prometheus alerts.
- COIR, a B3 brokerage-note extractor: pandas 2 and 3 compatibility fix (PR #34), incorporated into v2.2.0.
Platform and delivery: AWS, Kubernetes, Terraform, Argo CD, GitHub Actions, Docker, Ansible and Linux.
Observability and reliability: Prometheus, Grafana, Mimir, Loki, Tempo, OpenTelemetry, Alertmanager, Velero, Chaos Mesh and Trivy.
ML, data and FinOps: Python, PyTorch, scikit-learn, FastAPI, Bedrock, OpenSearch, PostgreSQL and Cost Explorer.
Postgraduate studies at Data Science Academy: Machine Learning and MLOps (finishing December 2026), and Mathematics and Applied Statistics for Data Science, ML and AI.
| Certification | Issuer |
|---|---|
| AWS DevOps Engineer Professional | Amazon Web Services |
| AWS ML Engineer Associate | Amazon Web Services |
| AWS Solutions Architect Associate | Amazon Web Services |
| AWS SysOps Administrator Associate | Amazon Web Services |
| AWS AI Practitioner | Amazon Web Services |
| FinOps Certified Practitioner | FinOps Foundation |
| NVIDIA AI Infrastructure | NVIDIA |
In progress: the Kubestronaut track (KCNA, CKA, CKAD, KCSA, CKS).
Happy to talk about any of these projects on LinkedIn.