I build infrastructure that survives crises I didn’t plan for.

What happened

In March 2026, drone strikes hit AWS me-central-1 (UAE). Our entire region went down. All databases and backup snapshots were rendered unrecoverable — not partially, gone. The mysqldump backups I had been running stopped working in the days that followed, so the primary data was not recoverable through the intended DR path.

What actually saved the company

The data team had an EC2-hosted Airbyte instance and AWS MWAA (managed Airflow) streaming data into Redshift — a pipeline I had provisioned and set up for them earlier. It was never designed as a DR mechanism, but it became the recovery path: data was exported from Redshift and imported into the AWS Jakarta region.

My role in the recovery

Over a ~2-week migration I owned the DevOps side: environment configuration, AWS Secrets Manager, SSM Parameter Store, deployment validation, keeping migration-phase deployments green, and cross-team coordination with data, backend, and QA.

Honest framing: I was not the person who “led the rebuild” or achieved “zero data loss.” Some customer data was corrupted and required manual verification and fixes alongside other teams. Backend and QA did the heavy lifting on feature correctness. I made the earlier infrastructure decision that turned out to be the lifeline, and I held the DevOps line for two weeks.

The lesson

Snapshot-based DR is fragile in a real disaster. The pipeline saved us almost by accident. Afterwards, we rebuilt the DR strategy around continuous replication paths we know work — not runbooks we hope work.

Why

For a cost-sensitive company, a paid observability SaaS is a recurring tax that grows with your data. At $815/month and climbing, New Relic was the obvious candidate to replace once the team had the Kubernetes muscle to run its own stack.

What I built

A self-hosted LGTM stack on Kubernetes: Loki (logs), Grafana (visualization), Tempo (traces), Mimir (metrics). Migrated all three signals off New Relic, then built unified dashboards covering application performance, logs, and deployments in one place.

Outcome

~50% cost reduction, better log retention, and more query flexibility than the managed product it replaced.

How 100% Spot works in production

Stateless services with graceful drain on interruption notice. Queue-backed async workloads with idempotent consumers. Autoscaling groups with mixed instance policies for diversification. And the rule that makes it all safe: zero stateful dependencies on Spot nodes.

The traffic burst — and scope discipline

During a Ramadan sale campaign, traffic spiked roughly 10x and AWS spend peaked at ~$66K/month for about two months. The root cause of the bottleneck was a Lambda authorizer — an application-layer problem the engineering team solved with bucket-based rate limiting.

My job was different, and I kept it that way: observability during the incident, on-call, and coordinating with AWS Enterprise Support to raise service limits and enable max scale-out — giving the team room to breathe while they built the fix. Knowing what’s yours and what isn’t is what keeps incident response fast.

Context

At a regulated digital-asset fintech, ISO certification isn’t a badge — it gates licensing. As the sole DevOps engineer, I was the primary internal owner for both ISO 27001 and ISO 27017, working with an external vendor consultant.

What I did

Defined and implemented controls across access management, change management, incident response, and infrastructure. Coordinated cross-functional work spanning engineering, security, legal, and ops, and drove the process through the external audit.

Outcome

Passed the external audit on the first attempt, unlocking regulatory licensing for the business.