Senior platform engineer · SRE · MLOps
I take over infrastructure nobody has mapped end to end, make it safe and observable, and turn it into code teams can ship on.
Five years on production platforms where downtime costs money: a stock-exchange trading system, a crypto exchange, a multi-tenant retail platform and a global reinsurer. Today I own infrastructure workstreams on the AI platform of a global mobility company, where I migrated an ML estate off self-managed Kubeflow onto Databricks and built the private DNS behind PrivateLink-only access. On the retail platform I led the DevOps team, and one migration I planned and ran cut a tenant’s running cost by about 80%.
What I work with
Everything here is on a project below.
- Cloud
- AWS (multi-account networking, PrivateLink, Route 53, Lambda, API Gateway, Step Functions, ECR), GCP, Azure
- Infrastructure as code
- Terraform, OpenTofu, Terragrunt, Atlantis, Ansible, CloudFormation
- Containers
- Kubernetes on EKS (clusters and workloads), GKE and AKS, Helm, Docker, Harbor
- CI/CD
- GitHub Actions, GitLab CI, Jenkins, Azure DevOps, release-please
- Reliability
- SLO and SLI design, Prometheus, Grafana, Alertmanager, CloudWatch, Datadog, PagerDuty, disaster-recovery design
- Security
- Cloud security audits, IAM and KMS, HashiCorp Vault, OIDC and workload identity federation, Keycloak, Trivy, distroless images
- ML platforms
- Databricks (Unity Catalog, Asset Bundles, cluster policies), MLflow, SageMaker, Kubeflow, Kafka, Spark Structured Streaming
- Data
- PostgreSQL (Patroni, HAProxy, Keepalived), MySQL, Delta Lake, BigQuery, Liquibase
- Languages
- Python, Bash, Go, TypeScript, Groovy
- AI-assisted work
- Claude Code with custom skills and sub-agents, MCP integrations with Jira, Slack and GitHub, automated pull-request and Terraform-plan review, local models through llama.cpp
Problems I take on
-
Nobody can draw your network anymore.
I map it, move control traffic onto private paths and keep the result in code. Last time: private hosted zones in every region and account, and a missing Cloud WAN return route nobody else had found.
-
Deploys work until one region doesn’t.
I bound every wait, kill the races and write recovery down with numbers. Last time: a deploy race that was losing version commits across eight regions, and an approved recovery design with a one-hour RTO target.
-
An auditor is coming.
I run the audit first, then carry the fixes through in code. Last time: an account allowlist checked against 3.19 million audit-log events and cut from 37 entries to 28, each with a named owner.
-
Your ML platform outgrew its foundations.
I build the governance, serving and migration groundwork data teams stand on. Last time: 51 pipelines and 1,500+ models moved off self-managed Kubeflow onto Databricks, and a quarter of the estate retired instead of moved.
Evidence
- Shipped built and in use
- Measured counted or benchmarked
- Found a risk discovered through evidence
- Target a design objective
- Plan proposed, not yet done
- 3.19M
audit-log events checked to cut an account IP allowlist from 37 entries to 28, each with a named owner. Details for 3.19M
Shipped - p99 91 ms
on a reproduced real-time feature path, with 99.4% of rows inside the 500 ms target. Production ran on the 8.1 s micro-batch path. Details for p99 91 ms
Measured - 4
platform defects exposed by my first end-to-end deploy from the Unity Catalog model registry, including model settings the MLflow deploy path dropped without an error. Details for 4
Found - 141
catalogs moved to a new data-governance group through reviewed Terragrunt changes, with import blocks bringing drifted resources under code. Details for 141
Shipped - 79
repositories checked for access before a tenant launch step could run. Details for 79
Shipped - 1 h
recovery time objective in the multi-region serving design. A design target, listed as one. Details for 1 h
Target - 1,500+
registered MLflow models, most of them per-city variants of one family, moved to Unity Catalog aliases in the migration. Details for 1,500+
Shipped - 80+
Dependabot vulnerability alerts closed across two ML monorepos. Details for 80+
Shipped
Selected work
-
Private DNS for a PrivateLink‑only data platform
The platform rule said no public control-plane endpoints. I built the private DNS that makes the rule work in every region and account, traced a missing return route, and cut the account allowlist to entries with owners.
-
Reliability for multi‑region model serving
A dual-registry image path forced by a platform limit, a deploy path for eight regions with no lost version commits, a recovery design with explicit targets, and the alerting chain behind it.
-
Retailer launches as plan‑first workflows
Onboarding changes for three retailers across 24 service repositories, an access module that finds missing repository access before a launch step runs, and a tenant migration that cut its running cost by about 80%.
-
A highly available PostgreSQL platform
Patroni for failover, HAProxy and Keepalived for one stable address, Ansible for repeatable nodes, and alerts on replication lag and failover events.
-
A migration built from a full inventory
A legacy Kubeflow estate, from 51 pipelines to more than 1,500 registered models, inventoried and scored before anything moved, then migrated to Databricks in waves.
How I work
-
Everything as code
One OpenTofu and Terragrunt repository, applied through Atlantis after review. Private networking, Databricks governance and the serving path all live there.
-
Read-only before write
I start with an inventory and an audit: what exists, what is exposed, who has access, what is patched. That first pass found an allowlist of 37 entries with no owners, and 51 Kubeflow pipelines nobody had listed.
-
Every change through review
Plans are visible before they apply. I review the platform’s Terraform as a merge gate, and my own changes go through the same gate.
-
Decisions in writing
I write the team’s ADRs on security, recovery and platform limits. The recovery design with a one-hour RTO target is one of them, and it was approved.



