Cloud infra built to last under load and attack.

Client running $30k+/month AWS bills with no autoscaling, single-region setup, and 4-hour release windows caused by fully manual deployments. Any outage meant total downtime — no failover, no recovery plan.

Rewrote the entire stack as Terraform modules, introduced multi-region failover via Route 53 health checks, replaced on-demand fleets with scheduled spot instances, and wired end-to-end CI/CD through GitHub Actions. Zero-downtime deploys from day one.

A multi-cloud environment with overly permissive IAM policies, secrets hardcoded in environment variables, and no audit trail. Enterprise customers demanding SOC 2 compliance before signing — the current setup would fail every control.

Designed zero trust architecture from scratch — least-privilege IAM rewrites across all services, HashiCorp Vault for secrets, GuardDuty + Security Hub for real-time detection, and a complete audit trail meeting SOC 2 Type II requirements.

50+ engineers blocked waiting for ops to provision environments. Deployments took days, no service visibility, and every team had invented their own snowflake process. Productivity was collapsing under manual ops toil.

Built a self-service internal developer platform on Kubernetes. GitOps via ArgoCD meant any engineer ships by merging a PR. Standardised Helm chart library, Prometheus + Grafana dashboards for every service, automated provisioning cut the ops queue to zero.

LLM-powered product running in a Jupyter notebook with 3–8 second response times, no batching, no evaluation, and completely unmonitored in production. Costs were unpredictable and regressions invisible.

Rebuilt the inference layer on FastAPI with async batching, a Pinecone-backed RAG pipeline, GPU autoscaling on EKS, and a full evaluation framework that catches quality regressions before they hit users. Latency dropped from seconds to milliseconds.