I build the services that connect enterprise platforms, and the tooling I use to ship them.

A platform that coordinates specialised AI agents through a governed six-phase lifecycle (refinement, specification, planning, execution, review and release) to turn an idea into released software. Every step produces versioned artifacts, people approve them at gates, and rollback and a full audit trail keep the state deterministic. Source available under the PolyForm Shield licence.

• Lifecycle state machine with strict, advisory or autonomous approval gates. Artifacts are versioned and hash-checked, with diffs, rollback and release versions. These business rules live in a pure, unit-tested Domain project behind thin ASP.NET Core minimal-API endpoints.
• Specialist agents behind one LLM abstraction (Anthropic, OpenAI, Ollama, plus a deterministic fake for tests). Code generation runs step by step, then opens pull or merge requests, runs review cycles, and merges and tags releases on GitHub, GitLab or Bitbucket.
• Composes a stack for each project from a versioned, content-hashed catalog, or infers one from an existing repo. Deploys to Vercel, Netlify, Cloudflare, GitHub Pages, Render or Railway with managed databases, and imports issues from Jira, Linear, GitHub and GitLab.
• One API process runs the endpoints and the background workers. Workers claim jobs from PostgreSQL with FOR UPDATE SKIP LOCKED, and a transactional outbox writes each state change and its follow-up jobs together, so there is no Redis or message broker.
• Next.js 16 frontend behind a custom Node server that proxies the API and SignalR WebSockets on the same origin and sets the CSP. It calls the API only through a TypeScript client generated from the OpenAPI contract. Notifications go out in-app and by email, Slack, Teams and signed webhooks, and there is also a tevrin CLI.
• GitHub Actions CI enforces warnings-as-errors and fails on OpenAPI contract or EF migration drift. Railway infrastructure is code: preview deploys after green CI on main, and production deploys on release. It also runs on Docker Desktop Kubernetes, and a Jekyll docs site is published to GitHub Pages.

A public, installable marketplace of Claude Code plugins for building apps from a spec on any combination of stack layers. One core plugin holds a general orchestrator and generic role-based agents, and these load the stack-specific skill packs that each project's STACK.md names. I used it to build Tevrin.

• stack-core: write-spec turns an idea, ticket or notes into SPEC.md. compose-stack picks one pack per layer, checks that they are compatible and writes STACK.md. orchestrate-stack builds in phases, with gates, reviews and commits, then runs the app locally and makes any deploys STACK.md permits.
• Ten generic agents (scaffolder, data, backend, frontend, QA, CI, deploy, reviewer, auditor and docs writer) that load the matching pack skills, so one set of agents works across every stack.
• 19 layer packs: React + Vite, Next.js and vanilla frontends; ASP.NET Core, FastAPI and Express backends; EF Core (PostgreSQL or SQL Server), SQLAlchemy and Supabase data; GitHub CI; deploys to Vercel, Netlify, Cloudflare, GitHub Pages, Render and Railway; and local Docker Desktop with or without Kubernetes.
• compose-stack knows which data packs fit which backend, what can deploy where, and what split deploys mean for CORS and cookies. An author-pack skill defines the interface a new pack must follow.
• Guardrails recorded in STACK.md: deploy autonomy (none, preview or production on request), no remote changes or data wipes without asking, and Kubernetes limited to the docker-desktop context.