The architecture: how the pieces actually fit together
Subatomic 101 series · Post 4 · Karl Simon, CTO
So far this series has built three things in order: the vocabulary (LLMs, agents, AI Co-Workers), the data foundation (bronze to silver to gold), and the cognition layer (your firm’s reasoning, codified on top of gold). This post is about assembly — how those pieces fit together into something that actually runs a firm.
I want to start with the problem the architecture exists to solve, because it’s the problem sitting inside almost every firm I talk to.
The Frankenstack
Walk into a typical firm and count the tools. A CRM. A custodian feed. A planning tool. A portfolio system. A document store. A scheduler. A compliance platform. An email system. A reporting tool. A few spreadsheets nobody will admit are load-bearing. You’re at 10+ tools before anyone has done any actual work, and every one of them was a reasonable purchase at the time.
I call it the Frankenstack. Each piece is fine. The assembly is the monster. The tools don’t agree with each other — we covered why in the data post — and the humans are the glue. Your best people spend their days copying information out of one system, reshaping it, and pasting it into another. Every new tool that promised to fix this added one more thing to reconcile.
Here’s the expensive mistake this post is about: when firms feel this pain, their instinct is to buy another app. Something with a better dashboard, one more integration, an AI feature bolted on. It never works, because the problem was never a missing tool. The problem is that nothing is orchestrating the tools you already have. You don’t need app number eleven. You need the layer that makes the other ten mean one thing and act like one system.
That layer is the architecture. Let me build it for you from the bottom up.
Start at gold
Everything sits on the gold data layer from post 2 — the single source of truth, where “Sam,” “Sammy,” and “Samuel” are one person and every number has one agreed answer. I won’t re-argue it. I’ll just repeat the rule: nothing above this layer is real until this layer exists. AI on bad data just makes faster mistakes.
Subatomic IQ: the cognition
On top of gold sits Subatomic IQ — the cognitive engine from post 3. The firm’s reasoning, decision-making, and SOPs, codified. This is the moat, and it’s the part that makes everything above it yours instead of generic. Gold tells the system what’s true. IQ tells it how your firm thinks about what’s true.
The Chief of Staff: the manager
Here’s where the architecture starts to look like an org chart instead of a tech stack, which is the point.
The Chief of Staff is the AI manager. One point of contact. You don’t learn ten interfaces and you don’t route requests yourself — you tell the Chief of Staff what you need, and it directs the workforce. It preps your day, dispatches the right specialist for the job, and brings the result back. A firm doesn’t want to manage a fleet of AI systems any more than a principal wants to personally manage every task in the building. You want one competent manager. That’s the role.
AI Co-Workers: the roles
Under the Chief of Staff sit the AI Co-Workers — the specialists. Meeting prep. Client follow-up. Compliance review. Data operations. Each one carries a role end to end, trained on your firm’s way of doing it, working inside the systems you already run. Not features. Roles. The unit here is a job description, not a capability list — which is why the right mental model for adopting this is hiring, not procurement.
Signal, Deep Lens, and Shield: the wrapper
Three more pieces sit around all of this.
Signal is how you interact with the workforce. It meets your team in the channels they already use — chat, email, voice, the CRM, whatever the firm actually lives in. Nobody learns a new interface. The tools stay. The behavior of the tools changes.
Deep Lens is the record of what happened. Every action a Co-Worker takes — what it accessed, what it decided, what it produced — is captured in a traceable audit trail. When compliance asks “who did this, when, and why,” the answer isn’t a reconstruction. It’s a record. In this industry, work you can’t audit is work you can’t defend, and Deep Lens makes every Co-Worker’s work fully inspectable by default.
Subatomic Shield is the security and governance boundary around everything: role-based access, so a Co-Worker sees exactly what its role permits and nothing more; governance rules the firm sets; and the firm’s data never leaves the firm. Deep Lens and Shield together are the trust layer — I’ll spend the whole next post there, so here I’ll just place them in the diagram: they aren’t features you turn on later. They’re the highly protected perimeter the rest of the architecture lives inside.
Orchestration is the unlock
Read the stack back from the bottom: gold data, IQ, Chief of Staff, Co-Workers, with Signal and Shield around it. Notice what isn’t in it: a new app for your team to learn. Notice what is: every layer exists to make the layers around it mean something. The data means one thing. The reasoning acts on the data. The manager directs the roles. The roles do the work inside your existing systems.
That’s orchestration, and it’s the whole unlock. The Frankenstack didn’t fail because the tools were bad. It failed because nothing conducted them. The architecture doesn’t replace your ten tools — it makes them finally act like one firm.
What changes for the humans
This is the part that matters to everyone who doesn’t care about architecture diagrams, so let me make it concrete.
Vantage Financial — a $1.5B AUM RIA — deployed the Subatomic Workforce and in year one saved 8,000 hours, unlocked $500K+ in operational value, and replaced the equivalent of 4 FTE across 4 departments. The number I want you to sit with isn’t any of those. It’s what Vantage did next: they didn’t cut those four roles. They redeployed those people to client-facing, revenue-generating work — the growth work humans should have been doing all along, back when they were busy being the glue in the Frankenstack.
That’s the trade the architecture offers. The machines carry the operational weight. The people go do the work only people can do. Growth without proportional headcount, and a team that stops being copy-paste infrastructure.
Next in the series: the question every regulated firm asks last and should ask first — how do you audit, govern, and actually trust any of this?