Slideware says it will work. Vendors say it will scale. Budgets get committed before anyone has measured anything. You can't govern what you never measured.
I start with the problem, not the tech. The build is never the point — somewhere there's a decision stuck for lack of evidence, and my job is to understand the business well enough to find it, then build whatever produces the proof.
AI projects, really IT projects, fail for organizational reasons, not technical ones. Most organizations don't have a structured way to think about those reasons before they become expensive lessons. A demo proves possibility. It does not prove the organization is ready.
I have always operated with a methodology; it just wasn't always formalized. SPARK is that methodology, written down — not another governance checklist to complete after the technology has already been selected. Anyone with a military briefing background may recognize the bones.
Everybody wants to be on the AI project until the project needs an owner.
Read the Anchor Article ↗A build discipline for AI-assisted engineering. AI writes code faster than anyone can verify it — I wrote the Forge Method to help close that gap, so what gets built is something you understand and a stranger can inherit, trust, and extend.
Green tests aren't proof. The AI's report isn't proof. A human reading the real output is.
These rules didn't come from a whiteboard — each one came from something a live build broke. Every build in the next section came through them; the full method follows SPARK.
Agent products live or die on unit economics. I wasn't willing to commit an architecture until I knew what a real session cost.
I built a game engine with an AI controller on Cloudflare's edge — Agents SDK, durable server-authoritative state, AI Gateway metering from the first call.
I exposed the shared capabilities as MCP services so the platform outlives the project.
Metered AI inference cost per 30-turn session: $0.097 — measured, not modeled — engineered to a $0.10 gate. Prompt restructuring cut uncached input tokens 89,315 → 20,743 (~77%).
I answered the cost question for about $3 in API spend — the price of a cup of coffee. At 10,000 sessions a month that comes to about $970, and I knew it before committing anything to production.
Basis: limited pilot · multiplayer unproven
Unit economics belong in the build, not the postmortem. Meter from the first call and you always know what the system costs to run. This is the diligence most AI vendors hope you never do.
The traffic looked wrong. But "looks wrong" doesn't justify blocking, escalation, or spend — suspicion is not a security decision.
The proof was buried in 45 million log lines no human would ever read.
I built an AI-assisted pipeline and turned it loose on the raw record — 90 days of production web traffic — hunting the addresses and patterns that separate an attack from noise.
45 million records, distilled to named addresses and attack patterns. The suspected activity was real.
Worse — it wasn't new. The analysis tied the current activity to infrastructure one network block away from a previously investigated incident. Same provider, same neighborhood, years apart.
Basis: infrastructure proximity and behavioral correlation · common-actor attribution unproven
Blocking decisions went from "we think" to "we know." Every public-facing site gets probed all day, every day. The real question is whether anyone would notice. This build did.
In a high-velocity inventory operation, changing strategy in production is a bet with real capital. The question: could strategy be tested in simulation first?
I built a simulation of the operation around its real operating data, so a strategy could be run over and over without touching inventory or capital.
Technical basis: deterministic hash-keyed retrieval (retrieval-grounded, no vector database) · Dijkstra shortest-path routing
The simulation surfaced a "third state" data problem — inventory in a state the system of record couldn't represent. Found in the model, not on the warehouse floor.
Strategies get tested before capital moves, and data-model flaws surface before they cost money. That's SPARK's first question answered in software: how does the work actually get done?
A lot of AI tools put the safety boundary in the prompt. I wanted a system I could trust with years of accumulated, high-value work — which meant the safety couldn't live there. It had to be architecture the AI physically couldn't route around.
I run a full AI production platform under house rules: every action routes through deterministic logic, not the model's own judgment. Retrieval is deliberate, not ambient — the system only sees what a task actually needs. And nothing destructive executes without a human confirming it first. The collaborator proposes. It never just acts.
Across sustained production use: zero unapproved actions. Destructive actions are blocked in the routing layer until a human confirms them. The model has no direct execution path around that control.
This is agent governance that actually runs in production, not a slide about agent governance. The question every board eventually asks — "who can stop the system?" — has one answer here: the human, every time. Because it was built that way from the start, not bolted on after something went wrong.
Enterprise assessments consume months of consulting hours on work that is largely structured. I decided to find out how much of that lift AI can carry while executive judgment stays where it belongs.
Recon asks the questions I would ask in an enterprise assessment, organizes the evidence, flags the contradictions, and maps organizational readiness — the groundwork a digital transformation roadmap gets built on.
Still in build, so the evidence is pending. This sheet gets its numbers before it ships — that is how the method works.
The goal is to shrink the cost of knowing where you stand — from months of consulting hours to weeks of working sessions. The structured lift moves to the machine. The judgment stays human.
Anyone can assert leadership, so this sheet holds to the same standard as the rest of the set. The evidence of leading well is what keeps running without you: the teams, the systems, and the people who took over.
A revision block means the same drawing, improved — never a new one. The names and dates are on my resume. This sheet is about what lasted.
Don't hire me to just keep the lights on. I keep the lights on by building systems and teams that don't need heroics. Hire me to take you further, when your business has outgrown its stack and integration, modernization, and organizational design all have to happen at once.
I've run technology where controls weren't optional, so governance is baked into how I build teams and systems. I lead so controls aren't just compliance. My orgs don't prepare for audits; they just get audited.
If everyone in the room is certain and nobody has measured anything, that's usually where I come in.