Foundation Operations · Mind

We don't sell software.
We build a brain for your business.

One company at a time. Locally hosted. Citation-grounded. Built after we've spent the time to actually understand how your operation works — not pulled off a shelf.

Engagement 12-week minimum
Concurrent clients Two. At most.
Inquiries / yr ~40 in, 2–3 selected
Status Accepting Q3 inquiries
01— Manifesto

There is no off-the-shelf brain.

A useful brain has to know your operation the way your most experienced employee knows it. That kind of knowledge can't be installed from a vendor. It has to be built.

The market is full of "AI for [your industry]" products. They're trained on the average of an industry and tuned for the median customer. They get rolled out, they make decent demos, and then six weeks later they're sitting in a Slack channel that nobody opens.

We do the opposite. We spend the first month listening — to your operators, your tickets, your edge cases, the war stories your supervisors tell new hires. Then we build a brain that knows your business specifically. Not your industry. Your business.

Every brain we build is custom. We never reuse a deployment. We never scale to ship. And we say no to most of the work that comes through the door.

02— The engagement

Twelve weeks. Four phases. One outcome.

Most "AI projects" fail in week three because nobody bothered to learn the actual job before writing the code. We don't write code in the first month. On purpose.

01
Weeks 1 – 4 · Discovery

We learn the business before we build anything.

30 days of structured listening. We sit in on calls. We read the playbooks. We shadow your top operators. We pull up the past year of tickets and look for the patterns nobody documented.

  • Operator interviews — your top three, ninety minutes each
  • Ticket & comms audit — twelve months back, sampled and read
  • SOP and training-doc inventory — what exists, what's stale, what's tribal
  • Write-up: the brief that everything else is built against
02
Weeks 4 – 6 · Design

We architect a brain that fits — not a brain that's generic.

From the discovery brief, we design two parallel implementations: one that touches no customer data and ships immediately, and one that touches everything and waits for compliance. Same intelligence, two perimeters.

  • Data inventory — what exists, what's reachable, what's regulated
  • Perimeter map — what runs where, who can reach what, what gets logged
  • Approval pathway — the conditions that unlock the regulated build
03
Weeks 6 – 12 · Build

We build it on hardware you own and can unplug.

The non-secure build goes into production. The secure build runs on synthetic data in parallel. Every answer comes with a citation. Nothing leaves your network. The system is yours, not a license you rent.

  • Local-first stack — no cloud LLM calls, no SaaS dependencies
  • Citation-grounded answers — every claim has receipts
  • Internal-network only — reachable from your team, no one else
  • Full audit log on the regulated side — who asked, what came back
04
Months 4 – 6 · Soak & handoff

We measure what changed, then hand you the keys.

60-day soak in real workflows. We measure hours saved, queue impact, escalation reduction. The brain isn't done because we shipped — it's done because your team is faster and your numbers moved.

  • Operational metrics — before, during, after
  • Operator survey — what works, what doesn't, what they'd change
  • Documented runbook + on-call escalation path for after we're gone
03— What we promise

Four guarantees. In writing.

These aren't marketing copy. They're the conditions we'll put on paper before signing a statement of work.

// 01

Local-first, always.

Every model weight, every byte of data, every working part of the AI runs on hardware you own. The "cloud" in this system is your office.

// 02

Citation-grounded, always.

Every answer comes with its receipts. If the brain can't point at a source, it won't pretend to know. No fabrications dressed up as policy.

// 03

Two builds, hard-isolated.

Customer data lives behind compliance, on its own server, with its own credentials. The useful build ships first. The regulated build waits for written approval.

// 04

Yours, not ours.

You own the code, the database, the deployment. We hand off documentation and an escalation path. No vendor lock-in. No surprise invoices.

04— Currently in production

A brain that knows the operation, not the customers.

A regulated records platform serving over a million customers needed a system that could answer agent questions about operations, vendor processes, and marketing — without putting a single byte of customer data in front of a cloud LLM.

Case study · 2026 · Records services

One million customers. Zero customer bytes in the model context.

Industry: Regulated records · Tenure: 12-week build, 60-day soak

The brain shipped in two builds. The open build went live in week four with the operations corpus — playbooks, vendor reference data, marketing intelligence, system runbooks. The locked build was fully constructed on synthetic data and parked behind compliance sign-off, ready to flip when leadership says go.

Same intelligence on both sides. Different perimeters. Zero shared infrastructure between them.

"If we can't say, by construction, that the open build couldn't return a customer name — we haven't built it right."
05— The two-builds principle

Why we never ship "one with a flag."

A flag is a footgun. A separate database, on a separate host, with a separate credential, means the answer to "could the agent-facing brain ever leak a customer name?" is no, by construction — not "no, if the flag is set right."

Build A · Open

Knows your operation. Knows nothing about your customers.

SOPs, training, vendor reference data, marketing intelligence, system runbooks. Customer information is removed at ingest — never lands here. Ships in week four.

Build B · Locked

The full brain. Gated on compliance.

Customer records, communications, case detail. Identical interface to Build A — agents call the same tools. Different perimeter, different audit trail, written sign-off required to unlock.

06— Fit

Who this is — and isn't — for.

We say no a lot. Not because we're precious about it, but because we only do this work well when the engagement is right.

Good fit if
  • You have an operation with real depth — playbooks, edge cases, tribal knowledge
  • Your data is regulated, sensitive, or otherwise can't be sent to a cloud LLM
  • You can give us access to operators and one month of structured listening
  • You want to own the system, not rent it
  • You can sign a 12-week minimum engagement
Not a fit if
  • You want a chatbot you can demo next week
  • You'd rather pay a SaaS fee than own infrastructure
  • You want to compete with us on architecture decisions during build
  • The data you'd put in front of it is messy and nobody wants to clean it
  • Compliance hasn't been brought into the conversation yet

Tell us about your operation.

We read every inquiry that comes in. If your situation is a fit, we'll set up a discovery call. If it isn't, we'll tell you that — quickly, and with a recommendation for what we'd do instead.

Start a conversation
Currently accepting inquiries for Q3 starts