I build the infrastructure between AI agents and production software — MCP and tool surfaces, repository intelligence, agent coordination, and the delivery pipelines that put them in front of real users.
Most of my public work sits in three places rather than here. This page routes you to them.
Prove It Resume — a technical evidence surface built so an evaluator can check claims instead of believing them. Every claim carries a stated boundary; every proof links to the artifact behind it. Six CI gates, including axe-core accessibility, run on every pull request.
Vreko — codebase and session intelligence for AI-assisted development. The core application is proprietary. The interfaces are public:
| Surface | Repository |
|---|---|
CLI (@vreko/cli) |
vreko-dev/vreko-cli |
MCP server (vreko-mcp-server) |
vreko-dev/mcp-server |
| VS Code extension | vreko-dev/vscode |
Organization: vreko-dev
workspace.json — an Apache-2.0 open standard for
committed repository intelligence, centred on the daemon-free .agents/workspace.json
artifact. I steward it.
It is an independent standard. It is not a Vreko feature and not a Marcelle Labs product.
| Repository | Owns |
|---|---|
workspacejson/standard |
Specification, JSON Schema, deterministic reference behaviour, governance |
workspacejson/cli |
Neutral producer (@workspacejson/cli) |
workspacejson/integrations |
MCP, Codex, CI and other host-side consumers |
Organization: workspacejson ·
Documentation: workspacejson.dev
Marcelle Labs — where the complete systems live: architecture through implementation, evaluation, deployment and verification.
- Interlock — evidence-bound coordination before shared-state mutation. A controlled counterfactual with a frozen evidence packet, an independent verifier, and a recorded Google Cloud traversal.
- Never Ask Twice — an enterprise support memory agent with a deterministic memory ON/OFF ablation, an MCP surface, and a live Qwen-backed deployment.
Organization: Marcelle-Labs
I prefer small, compatibility-first contributions that solve a maintainer's existing problem before proposing a broader integration.
A recent example: an additive workspace.json lookup change for
Buildomator. It shipped in v3.4.6
as 41459a5,
cherry-picked with authorship preserved and guarded by a regression test — the
maintainer's
confirmation
is on the pull request.
No peer-reviewed publication or preprint yet. What exists publicly:
- Decision-time repository evidence — which repository-derived signals, available at the moment an agent decides, causally change its plan. Bounded evidence in Interlock and in the Tally case study.
- Tool-selection and routing overhead —
vreko-dev/sopr-mcpdocuments the Service-Oriented Protocol Router pattern as a versioned artifact.
Agents should not have to rediscover the same repository facts every session, and safety-critical decisions should be grounded in inspectable evidence rather than prompt folklore.





