AI Operations, One Founder: What Scaling a Relay Harness Actually Looks Like

Three years ago, if you told a CPA or M&A advisor that a solo founder could assemble an AI operations platform with specialist roles, 214+ console API routes, configured Stripe billing, and 2,200+ passing tests across the workspace, they’d ask how large your engineering team was.

The answer is zero. It’s me, a veteran founder with an RTX 4060 and a multi-model AI development harness.

This isn’t a pitch. It’s a build report. Here’s what scaling an AI operations platform actually looks like when you’re doing it alone, and why the results matter for professional services firms evaluating AI investments.

The Numbers Are Real. Verify Them.

Let me start with what exists today, not a roadmap slide:

• AI specialist roles organized across multiple business functions
• Multiple council modes where agents debate decisions with structured dissent
• 2,200+ passing tests across the workspace
• 214+ API routes in the console (Relay Deck)
• 134 Supabase migrations in the console history
• Stripe billing integration configured for live subscription capture once production price IDs and webhooks are activated
• 3 published npm packages (open-source CouncilVerse framework)

Everything runs on Cloudflare Workers, Supabase, and Vercel. No vapor. No “coming soon.” The git history is the audit trail.

Council Architecture: Specialists That Disagree

Most AI platforms work like this: you send a prompt, you get a response, you trust it or you don’t. That’s fine for drafting emails. It’s dangerous for business decisions.

Our AI specialists don’t just answer questions. They argue about them.

The council architecture organizes 51 specialists into multiple council modes, each with a distinct decision methodology:

• Leadership Council debates like a board of directors with productive internal tension
• Founders Board strips decisions to first principles
• Dark Forest maps competitive threats and blind spots you’re not tracking
• Strategy Room pressure-tests under time constraints
• The Crucible stress-tests assumptions until they break or hold

When you ask a question like “Should we acquire this practice?” or “Is this pricing defensible?”, you don’t get one answer. You get a structured debate with a majority recommendation, a minority report, and a dissent score.

Every council member votes. Every vote is recorded. Every disagreement is preserved.

Why Auditable Dissent Matters

Here’s the problem with most AI advisory tools: they give you confidence without accountability. A single-model response feels authoritative. It reads well. But you have no way to know what it didn’t consider, what counterarguments exist, or how fragile the recommendation is.

Auditable dissent fixes this. When three agents vote against the majority recommendation, you can read their reasoning. When an agent flags a risk that the consensus dismissed, that’s in the record. When you make the final call, you made it with full visibility into the disagreement, not just the winning argument.

For CPAs reviewing an acquisition target, this is the difference between “our AI says it’s a good deal” and “our AI debated it across six dimensions, four agents flagged the revenue concentration risk, and here’s the minority report.” One of those is defensible to a client. The other is a liability.

We call this category Auditable Intelligence: where AI experts disagree, you decide.

Open Source: CouncilVerse on npm

The council engine isn’t locked inside our platform. The core framework is open-source, published as three npm packages under the CouncilVerse name:

• @councilverse/formations (v1.0.1): council structure and agent organization
• @councilverse/voting (v1.0.0): voting protocols and dissent tracking
• create-councilverse (v1.0.1): scaffolding tool to spin up your own councils

Anyone can install these, build councils, and run structured debates. We built the platform on top of CouncilVerse, but the framework stands alone.

Why open-source the core? Because trust in AI advisory requires transparency. If a firm is going to rely on AI-driven analysis for client decisions, they need to understand how the sausage is made. Black boxes don’t survive due diligence.

A2A Protocol v2.2: Not a Walled Garden

The platform implements A2A (Agent-to-Agent) Protocol v2.2 compatibility. In plain terms: our agents can communicate with agents built on other platforms using a shared standard.

This matters for enterprise buyers who don’t want vendor lock-in. If you build workflows on RelayLaunch and later want to integrate agents from another A2A-compatible system, you can. Your investment in council configurations and agent specializations isn’t trapped.

For firms running multiple AI tools, interoperability isn’t a feature. It’s a requirement.

BYOK: You Control Your Costs

Every AI platform has a cost problem. Usage-based pricing means your bill scales with adoption, which creates a perverse incentive to use the tool less as it becomes more valuable.

We solve this with BYOK (Bring Your Own Key). Customers connect their own API keys from providers like Anthropic, DeepSeek, or others. The model calls hit their accounts directly. They see exactly what they’re spending, they control their rate limits, and they’re never locked into our markup on tokens.

The platform charges a flat subscription for the infrastructure, the councils, and the tooling. The AI compute is yours to manage. For a CPA who bills $400/hour and uses the platform to run a structured analysis in 10 minutes instead of 45, the model cost per query is measured in cents. The ROI math is obvious.

What’s Actually Possible Now vs. Three Years Ago

In 2023, this would have required 8-12 engineers, a $2M seed round, and 18 months. I know because I’ve spent eight years inside enterprise operations. I know what headcount these projects used to demand.

Today, a solo founder with the right tools can ship production software at a pace that looks impossible by legacy standards. Not because AI writes perfect code (it doesn’t), but because the iteration cycle collapsed. What used to be a two-week sprint is now a two-hour session.

The test suite exists because when you’re the only engineer, you need the machine to catch what a code reviewer would. The API surface exists because real users need real functionality, not a demo with three endpoints.

This is the new math: one person plus AI can build what used to require a funded startup. Not everything. Not yet. But more than most people realize.

What This Means for Professional Services

If you’re a CPA, an M&A advisor, or a consulting partner evaluating AI tools, here’s what I’d want you to take away:

Look for auditable outputs. Any tool can generate a recommendation. Few can show you the structured disagreement behind it. If you’re advising clients on high-stakes decisions, you need the minority report, not just the majority opinion.

Look for open architectures. If you can’t inspect how the system reaches conclusions, you can’t defend those conclusions to clients, regulators, or in litigation. Open-source components and standard protocols are table stakes for professional credibility.

Look for cost transparency. If you don’t know what each AI query costs, you can’t model the unit economics of offering AI-enhanced advisory services. BYOK isn’t a nice-to-have. It’s how you build a sustainable practice.

Look at what one person can build. Not because solo founders are the future of enterprise software, but because the capability curve has shifted. If one veteran with a graphics card can ship this, imagine what your firm could do with a focused AI operations investment.

The question isn’t whether AI will change professional services. It’s whether you’ll be the firm that figured it out early, or the one that’s still evaluating vendors when your competitors are already running councils.

We built the platform. The agents are debating. The tests are passing. The npm packages are published.

Your move.