DJ Von FrankAI Implementation
AI / LLM Implementation
I build the AI systems companies actually use.
Your company bought access to AI. What followed was a collection of licenses and experiments, but no durable new capability. That is not primarily a model problem. I install the Company OS: the knowledge, governance, workflows, and adoption layer that makes AI part of how the company operates.
Seven years working inside and leading marketing teams taught me that implementation only matters when people adopt it. That is why adoption is designed into every system from the start.
The Build
Context architecture, routing, guardrails, and quality gates that stop flawed work before it reaches production.
4 company brains running in production
See the brains →The Adoption
Onboarding, training, and the unglamorous operational work that turns a clever system into something more than 100 people rely on every day.
100+ people inside one of them, every day
How that works →TRY IT / THE SAME REQUEST, TWO SYSTEMS
Five things a bare model cannot do, however well you prompt it: write in the company’s voice from the company’s facts, refuse when it does not know, check before it ships, change one fact everywhere, and remember what was decided.
A bare chatbot
Subject: Welcome to the Apex family!
Dear Valued Partner,
We are thrilled to welcome you aboard! At Apex Instruments we pride ourselves on award-winning quality and an industry-leading 48-hour turnaround. As a new partner you will enjoy 20% off your first order and access to our dedicated 24/7 support line.
We can’t wait to exceed your expectations! Reply to this email or call 1-800-APEX-LAB anytime.
No tone chosen. Four promises the company never made.
Fluent, warm, and ready to send. Also wrong.
Through the Company OS
Routed · tone locked · promises quoted from the one file allowed to make them · template instantiated, not freelanced
Subject: Your lab, your rep, your terms in writing
Dr. Ridgway, nothing about how your cases move changes this month. Your account manager stays your account manager, and your pricing is locked for six months from today.
In writing, so you can hold us to it: a six-month rate lock, a six-month Continuity Guarantee on turnaround, a $500 digital-integration credit, and twelve months of complimentary continuing education for your team.
If anything in your first month feels different from before, tell your rep. That is what the guarantee is for.
Every promise traces to pricing.md, the written Day-1 terms. No figure invented. No em dashes. One tone, never blended.
A bare chatbot
The Meridian bench typically ships in 10 to 14 business days.
No file opened. No source named.
Plausible, fluent, and unverifiable.
Through the Company OS
Opened · no Meridian bench entry
UNKNOWN
Meridian bench lead time
Owner
Marcus
Due
2026-09-12
filed to questions/register.md
A bare chatbot
Prices rise 3 percent on October 1.
One step, no gate.
The wrong figure ships.
Through the Company OS
The gate
↑ Back to build · two failures listed
A bare chatbot
It edits the document in front of it. The other two copies are never opened.
Through the Company OS
3 percent2.4 percent
Changed once · 2026-08-14 sheet
Three documents regenerate from the one fact.
A bare chatbot
I do not have information about your company’s decisions.
Nothing invented here.
Nothing remembered either.
Through the Company OS
Tiered discounts 5 / 8 / 10 percent
Dotted underlines mark what the generic model invented. Mono source lines name the file each grounded statement came from.
The Argument
I spent six years watching organizations buy tools that never became part of the work. AI is often treated the same way: buy the software, tell everyone to use it, and mistake experimentation for transformation. The model may be capable. The surrounding system is missing.
Failure one
What I build against it
Failure two
What I build against it
I noticed these failures because I came to AI through marketing. Both land on the desk of the person responsible for sending the work. For years, that was me. I reread every line because I could not tell what the tool had verified and what it had invented. The model was only one part of the system. Everything around it was missing.
The Four Layers
Most AI efforts overinvest in one layer and stall before the others connect. The value comes from making all four work together.
The part most companies think is the whole job, and the smallest share of it. Model access, the routing that decides which files a request opens, the integrations into the tools the work already lives in, and the pipeline all of it ships through. Every piece of it is inspectable, which is why a wrong answer here becomes a file someone in the company edits, not a ticket filed with a vendor. I build with LLMs. I do not build the underlying models.
The company’s actual facts, structured so a model cites them instead of inventing them. This is the Company Brain, and it is the layer that decides whether anything the system produces can be trusted enough to send.
What the system is actually pointed at. A chatbot that answers questions about the company demos beautifully and dies at single-digit adoption, because answering questions is nobody’s job. Producing the Tuesday deliverable is.
The half that decides whether the rest was worth building. 20+ people onboarded one at a time, and 100+ people opening the result every day. Rollouts do not die of bad technology. They die of silence.
Connected, the four stop being projects and start being a way of operating. Approved work returns to the brain as the reference the next request builds on, so the system is worth more in month twelve than it was in month one.
The test is one question: what would remain if the AI subscription were cancelled tomorrow? A company that experimented loses some personal conveniences. A company that operates still owns the structured knowledge, the written rules, the gated workflows and the trained habits, and points all of it at whatever model comes next.
The four layers of the Company OS. Open any one of them. The connection is the work.
The Operating Loop
Every request moves through the same seven-stage path. Its location in the folder tree shows its approval status, so nothing ships by accident.
Anyone can ask in plain language, for a deliverable, a fact, or a plan.
Incomplete requests trigger focused multiple-choice questions instead of guesses.
The constitution opens only the files the task requires.
Every governed fact has one authoritative source. Missing information is marked UNKNOWN and becomes a tracked question.
Skills use a gold-standard reference as the structural template, then adapt the content.
Automated checks enforce banned terms, tripwires, and house style. Pass or fail, not an opinion.
Only a human can move work to approved. Every previous version stays recoverable.
UNDERSTAND / THE COMPANY OS IN TWENTY SECONDS
Then every request runs the same loop: one failed check, one revision, one approval, and the result feeds the next request.
THE WORK / EVERY ONE OF THESE OPENS
Three families. The first is what the system does inside a company, the second is what it hands people, and the third is what is being built next. Every sample uses a company that does not exist.
CAPABILITIES 30
Checkers that read the work, a deck generator, a price rebuild that verifies its own output.
COLLATERAL 12
Flyers, catalogs, guides, price sheets, forms, email, social, direct mail, and two browser tools.
NEW CAPABILITIES 6
Ask the sample brain and get citations. Score the skills. Model the return on an engagement.
LEARN / NINE COMPONENTS, ONE DRAWING SYSTEM
TAKE SOMETHING WITH YOU
PDF · ONE PAGEFifteen terms, plain definitions, one printable page.
Download the PDF →
PDF · TEN SLIDESThe argument, the loop, the engagement, and what you own at the end.
Download the deck →INTERACTIVE · PRINTABLENine checks for AI guardrails that actually hold, with a printable version.
Open the checklist →
OPEN THE HOOD
Browse a sample company’s knowledge, brand system, and working files. Follow one request through four stages, each opening a real document.