Skip to content

I turn your team's business expertise into AI workflows that ship — and refuse to guess.

Senior AI Product Owner & Agile Project Manager — Scrum / SAFe, 20 years. Not just AI agents: a method that turns real business expertise into executable, reusable AI processes — proven live end to end, and the failures are published too.

See how it works Try it yourself → Talk to Guy →

Most AI projects fail on the question, not the model.

A stakeholder sends an incomplete request. The AI fills in the gaps, produces a confident answer, and gets it wrong.

I build controlled AI systems that refuse to guess.

When a request is incomplete, the system identifies what is missing and asks the right questions in plain business language. Each step is validated before the next one runs, and every execution is recorded.

The result is auditable work built on complete inputs — not confident answers based on assumptions.

38Versioned Agents
37Skill Folders
10Executable Workflows
9End-to-End Proofs

Those 38 versioned agents and 37 skill folders make up the catalog — every skill mapped to the business request it answers, and it's public. Browse the repository →

The system at a glance

One integrated system. Start with a business request: a free-form brief can be handled by the Dispatcher, which identifies the appropriate process, asks for missing information, or refuses the request when no suitable process exists — rather than letting AI guess.

flowchart LR
  A["01 · EXPERTISE<br/>Catalog"] --> B["02 · PROCESS<br/>Workflows"]
  B --> C["03 · CONTROL<br/>Agentic Runtime"]
  C --> D["04 · RESULT<br/>Auditable result"]
flowchart TB
  A["01 · EXPERTISE<br/>Catalog"] --> B["02 · PROCESS<br/>Workflows"]
  B --> C["03 · CONTROL<br/>Agentic Runtime"]
  C --> D["04 · RESULT<br/>Auditable result"]

From business expertise → to a structured process → to controlled AI execution → to a result you can check: passed, failed, missing information, or returned for rework. See how it fits together →

Proof at a glance

Nine workflows proven live, end to end — plus one returned for rework by its own quality gate. All ten have been through a live run, and the record includes the one that didn't pass:

  • ✓ WF-001 — 3 / 3 steps passed
  • ✓ WF-002 — 5 / 5 steps passed
  • ✓ WF-003 — 7 / 7 steps passed
  • ✓ WF-004 — 6 / 6 steps passed
  • ✓ WF-005 — 3 / 3 steps passed
  • ✓ WF-006 — 6 / 6 steps passed
  • ✓ WF-007 — 4 / 4 steps passed
  • ⟲ WF-008 — returned for rework: the gate withheld the report
  • ✓ WF-009 — 6 / 6 steps passed (provided candidate pool)
  • ✓ WF-010 — 4 / 4 steps passed

A record that only contains the runs that worked is a demo.

The system refuses to start on a request too thin to work from — handing it back with the missing questions named in plain business terms — and routes the rest to the right process, nineteen of twenty test briefs. See the dispatcher runs.

Per-step verdicts captured as JSON in the repository. See Live proofs.

Explore

  • The Stack

    The catalog, workflows and runtime as one system — 38 versioned agents and 37 skill folders, orchestrated into 10 workflows and executed by a gated runtime. Proven live, end to end.

    Explore the stack →

  • Projects

    Products and case studies built with the stack — starting with RAWLY, an evidence-first nutrition app where content is code.

    See the projects →

  • About

    AI Product Owner / Agile Project Manager — 20 years. Business analysis (MOA) and web integration end to end.

    About me →

Catalog currently at v4.4.0 · Agentic Runtime at v0.18.0 — both versioned, with sources in their repositories.