Nine deep dives. Each one started as a real problem — usually one that dozens of teams or a hundred clusters had at the same time — and ended as a system that solved it without a human in the loop.
A Kubernetes platform that runs large-scale LLM batch inference on interruptible GPU capacity. Custom scheduling policies, checkpoint/retry fault tolerance, and Karpenter elasticity treat Spot reclaims as normal events — sustaining ~80% utilization across 500+ GPUs while cutting compute cost by ~60%.
One Terraform module that provisions a complete microservice platform: shared ECS cluster, ALB host/path routing, blue/green external services, service-discovery internal services, and an independent CodePipeline per microservice. Teams write ~50 lines of HCL and get zero-downtime deploys, automatic rollbacks, and a versioned, audited foundation.
Kubernetes upgrades are scary on one cluster. This is a 4-phase orchestrator that does them across a 150+ cluster fleet: control plane, node groups, addons, then automated sanity checks — concurrent, cross-account via SAML/STS, resumable after partial failure, with a CSV audit trail for every run.
An ArgoCD app-of-apps framework where ApplicationSets discover clusters by label and a per-cluster Git inventory decides which addons each cluster gets — external-secrets, Traefik, OTEL collectors, security agents. Enabling an addon on any cluster in the fleet is a one-line Git commit; drift self-heals.
Three tools, one goal: know the state of everything. Concurrent health polling of hundreds of clusters, async endpoint checks for thousands of services with aiohttp + Prometheus metrics, and post-change validation that catches regressions before users do.
500+ CI/CD pipelines across the fleet needed the same upgrade. Instead of hundreds of tickets: a ServiceCatalog-driven engine that discovers, upgrades, and verifies every pipeline — idempotent, multi-region, ring-based rollout, with CSV audit trails per account.
A platform-team-in-a-box: REST API + React UI that takes a team from "we need infra" to onboarded — generating convention-correct Terraform, automating Git PRs, wiring notifications. Plus containerized multi-tenant app patterns with Node and Flask.
An agent designed like an SRE works: plan the investigation, fan out MCP tools in parallel, critique its own conclusion against live cluster state, and answer only when evidence converges. Langfuse-traced and eval-gated, RBAC-scoped reads, writes as GitOps PRs behind human approval.
A Helm-deployed monitor that discovers every ingress and service in its cluster and probes them concurrently with asyncio + aiohttp — because "Running" doesn't mean "up." Results export as Prometheus metrics; part of the platform behind 2,500+ monitored endpoints.
Each links to a deep dive — or I'm happy to walk through architecture decisions live.