One engineer's platform org, run as a playground.
Multi-agent tooling, infra-as-code, and MCP servers. Claude + Codex + Gemini in parallel.
- stark-skills - multi-agent PR code review. 3 AI agents × N domains, hierarchical config, prompts that improve themselves. The flagship.
- stark-marketplace - share skills, prompts, commands, agents, and MCP servers across
Claude Code,Codex, andGemini. - agent-native - my fork of Builder.io's framework; my GCP deployment of it (Cloud Run behind one origin,
Terraform+ Actions, Cloud SQL on a private IP) runs privately.
35 repos, most of them private. What that adds up to, solo:
- Multi-cloud
Terraformfoundations that provision and validate themselves - Workload Identity Federation, split plan/apply CI identities,KMS, a registry that drives the subnets, one shared LB + wildcard cert. - A vendor-admin layer over a dozen SaaS APIs, exposed as one
MCPtool across 400+ capabilities. - Self-hosted observability - logs, metrics, dashboards, alerts - on its own tier.
- A Slack knowledge server:
BigQueryhybrid search with an agentic ask layer. - A transcription pipeline -
Speech-to-Text v2into multi-stage LLM enhancement. - Image and video generation behind
MCP-Veo, animated media, brand icons. - A second brain, with its own
MCPserver and CLI.
- Branch + PR for everything - no exceptions
- Every review's findings land on the PR - nothing lost
- Test live - the real cloud surface, not localhost
- Go / TypeScript - no new Python
Ships to
main. No SLAs. That's the point.