Home Work
Work, written up properly.
Each case study covers the situation, what I built, the decisions behind it and what changed. Client details are kept general, but the approach and the results are real.
State Farm ML platform
2025 to now. Multi-cloud infrastructure for training, deploying and running machine learning models on AWS and GCP.
40% shorter release cycles
Faster path to production
One delivery path for ML services across GitHub Actions, Jenkins and Buildkite, with infrastructure from code.
GitHub Actions, Jenkins, Buildkite, Terraform, CloudFormation
95% fewer release interruptions
Releases nobody notices
Blue-green and canary rollouts on EKS and GKE, driven by GitOps, so new versions reach users in small steps.
Kubernetes, Helm, ArgoCD, FluxCD
30% faster training, 15% lower cost
GPU platform for ML
Autoscaling GPU and TPU clusters, plus SageMaker, Bedrock and Hugging Face wired into deploys.
GPU and TPU clusters, Kubernetes, SageMaker, Bedrock, Hugging Face
25% less downtime
Seeing problems first
Metrics, logs and dashboards in one place, alerts that mean something, and honest root cause reviews.
Prometheus, Grafana, CloudWatch, ELK
3 secret managers, one approach
Security in the pipeline
Secrets out of code, and scans that fail builds instead of filling reports.
Vault, Secrets Manager, Trivy, hadolint, Snyk, SonarQube
My own projects
Built in public, so you can read every line and every pipeline run.
$0 a month to run
This website
A hand-written site that ships like production software: checks, scans, a container, scheduled health checks and a live uptime panel.
GitHub Actions, Docker, nginx, Trivy, Python, GitHub Pages
One command starts the whole stack
A monitoring stack with tested alerts
Recording rules, an error-budget burn alert and a dashboard defined as code. CI breaks the demo service on purpose and checks that the alert fires.
Prometheus, Alertmanager, Grafana, Docker Compose, Python, promtool
20% → 60% → 100% canary, tested on a real cluster
A canary release that stops itself
A Helm chart with Argo Rollouts and ArgoCD. CI ships a good release and a broken one, and checks the broken one never reaches users.
Helm, Argo Rollouts, ArgoCD, kind, GitHub Actions, kubeconform, Trivy
Zero GPU nodes until a job needs one
A GPU cluster that costs nothing when idle
An EKS cluster with a tainted GPU node group on Spot. API access is limited to listed IPs, and CI scans every push.
Terraform, AWS EKS, VPC, NVIDIA GPU nodes, Spot instances, tflint, Trivy
Earlier work
Smaller wins from before State Farm. The full picture is on the experience page.
- Half the manual provisioning. Reusable Terraform modules for S3, IAM and EC2 at Amistad Technologies.
- 10% lower AWS bill. Found idle resources and scheduled automatic shutdowns.
- 20% fewer failed deploys. Checked infrastructure readiness during release windows.
- 95% SLA adherence. L1/L2 support for EC2, S3, IAM and networking incidents.
- Less babysitting. Python and Bash scripts with Lambda for backups, log archiving and IAM provisioning.
Hiring for DevOps or platform work?
I’m open to DevOps, SRE and platform engineering roles. Email is the fastest way to reach me.