Real work, in production today.

A development agency with its own product portfolio and clients in auto parts, HR, inbound tourism, healthcare and education. Every account arrived with its own stack — Go and Node APIs, Next and Astro frontends — and everything ended up on shared cPanel hosting, with no repeatable path to production and no way to reproduce an environment.

Two platforms, each for what it does best: a Docker Swarm with Caddy and automatic TLS for most services, and a k3s cluster for what needs high availability — PostgreSQL HA with CloudNativePG, Valkey with one primary, two replicas and three Sentinels, with Envoy Gateway and MetalLB in front. Deploys run on GitHub Actions: build, push to ghcr.io and docker stack deploy over SSH, the same across all 14 repos. Secrets are read from 1Password at provisioning time; none live in plain text.

69 application services in production across Swarm and Kubernetes. The main node has been running for 479 days without a single restart. A change on main reaches production in a 3-minute median, and 99% of deploys finish in under 15.

Nuke AI builds product for several clients at once: tourist transportation operations, hotel competitive intelligence, biomedical research and a dozen brand sites. Thirteen repositories, four languages and delivery cycles that look nothing alike — a static landing page and a live operations platform don’t ask for the same thing. The demand was never traffic: it was getting all of that to live on a single infrastructure without one client being able to take down another, and letting any of them ship to production the same day.

A Docker Swarm with Caddy in front handling automatic TLS per domain, and 23 isolated stacks on top: APIs, frontends, two PostgreSQL 16 instances, a pgvector database for embeddings, Redis and two n8n instances running the clients’ automations. Every project comes in the same way — GitHub Actions builds, publishes to GHCR and deploys over SSH, waiting for the stack to converge, verifying the SHA live, with a rollback workflow to a previous commit. The operations platform, NestJS and Next.js on PostgreSQL, runs a real-time war room with Socket.io over Valkey, Meilisearch search, write-invalidated caching and a 400 requests-per-minute-per-IP limit at the edge. The AI pipelines run on Claude with local embeddings and one hard rule: if a data point couldn’t be obtained, it is reported empty; it is never filled in.

23 services in production for more than ten clients on a single node, with no replica down. 423 days of continuous operation without a restart. A change on main reaches production in a 3-minute median, and 96% of deploys finish in under five.