Problem, constraint, build, number.
// the case studies
One case per system. Every number on this page traces to a receipt.
01GitOps at 100+ clusters
100+ clusters · 58 applications
- PROBLEM
- A hundred-plus cluster fleet fails by drift: hand-edited configs, snowflake clusters, no way to prove what shipped where.
- CONSTRAINT
- Classified environments, so promotion needs an audit trail and rollback cannot depend on a support call.
- WHAT I BUILT
- Declarative GitOps for 58 applications with Argo CD: promotion between environments, automated rollback, full artifact traceability.
02Release automation
~80% less rework · provisioning weeks → hours
- PROBLEM
- Provisioning cycles ran weeks, and deployment rework ate the platform team: the same fixes, reapplied by hand, cluster after cluster.
- CONSTRAINT
- The automation runs where it cannot phone home, and every change has to be provable after the fact.
- WHAT I BUILT
- Release automation in Python across the fleet: gated waves, health checks, rollback points, structured machine-readable results.
03Air-gapped LLM inference platform
100+ GPUs · air-gapped inference
- PROBLEM
- Teams needed dependable model inference in an air-gapped environment without a cloud fallback.
- CONSTRAINT
- Every dependency, deployment path, and recovery path had to be owned inside the environment.
- WHAT I BUILT
- LLM inference on a 100+ GPU fleet, including bare-metal and virtualized scheduling, storage, and gateway routing, run with the same release discipline as the rest of the platform.
04AgentOps
400+ stars · 50+ skills
- PROBLEM
- Agent output is cheap, confident, and wrong somewhere. Review by feel is how that wrongness ships.
- CONSTRAINT
- Runtime-neutral by design. The context that wrote the candidate never issues its binding verdict.
- WHAT I BUILT
- An open-source operating loop in Go: one intent, one bounded build, one fresh verdict, and 50+ portable skills. The result binds acceptance, changed files, and check evidence in an inspectable record.
05Gastown Operator
3 ecosystem projects: Gastown · Beads · KAgent
- PROBLEM
- Multi-agent work needs a control plane: workspaces, coordination, lifecycle. Kubernetes already knows how to run that shape.
- CONSTRAINT
- Operator discipline, so state lives in CRDs and a reconcile loop, never in bespoke daemon glue.
- WHAT I BUILT
- A Kubernetes operator in Go with controller-runtime: CRDs for multi-agent workspace orchestration, alongside ecosystem contributions to Gastown, Beads, and KAgent.
06Air-gapped AI platform (reference)
139 Helm templates
- PROBLEM
- Standing up an AI stack in a disconnected environment means re-deriving the whole map every time: inference, vector stores, gateways, observability, agents.
- CONSTRAINT
- Installs from vendored artifacts only, deploys by GitOps, nothing implicit.
- WHAT I BUILT
- A full air-gapped AI platform reference: the complete inference and agent stack, templated end to end.
// deep dives & open source
The pattern pages and public repos behind the cases.
- Platform ToolingAutomation for upgrades, diagnostics, secrets, backups, recovery, and day-2 operations where reliability has to be inspectable.
- OpenShift PaaSEnterprise container-platform patterns for restricted self-managed environments: multi-tenant, GPU-aware, GitOps-driven.
- Dify Agent PlatformA self-hosted AI application platform adapted for restricted OpenShift with observability, repeatable deployment, and operator-safe handoff.
- AgentOpsCoding-agent plugin for any model: Claude, GPT, and more. Go CLI plus skills for memory, validation, tracked work, and reusable operating loops.
- 12-Factor AgentOpsOperational principles for making AI agent work reviewable, testable, and safe to ship.
- vibe-checkA validation CLI that reads git history and turns AI-assisted coding sessions into reliability signals. Its ideas grew into the validation gates in AgentOps.
Every build here comes off one repeatable loop.
The cases are receipts. /method is the loop that produced them.