Stated Plainly

On the unit

The correct unit of AI deployment is the individual practitioner, not the institution.

Institutions move on procurement timelines. By the time a committee has selected a vendor, the underlying technology has shifted twice, and the selection is two generations behind the frontier. Institutions buy for the median user, which means the practitioner with thirty years of specific expertise gets a tool calibrated to someone with three. Institutions optimize for vendor accountability, which is the opposite of optimizing for outcomes.

The individual practitioner does not have these constraints. A solo attorney can adopt a workflow on a Tuesday and be using it Wednesday. A clinician with a clear sense of the friction in her own day can specify exactly what she needs and recognize whether she got it. A financial advisor knows the difference between a tool that reads his client and a tool that erases what he understands about that client. Decisions made at the unit of one expert, by that expert, on that expert's own infrastructure, produce outcomes institutions cannot reach.

This is not a romantic position. It is structural. The friction that prevents AI from delivering on its promise is the friction between what the technology can do and what a specific person needs it to do. That friction is highest at the institutional level and lowest at the level of one practitioner who knows their own work.

I work at the lowest-friction level on purpose.

On what AI is for

AI is for the augmentation of expertise that already exists. It is not for the replacement of it.

The version of AI that gets press is the version that promises to replace experts — automate the lawyer, automate the doctor, automate the financial advisor. That version is downstream of a misunderstanding. The thing those experts do that takes thirty years to build is judgment, not throughput. AI generates throughput cheaply. Judgment is what tells you which throughput matters.

A correctly built AI system gives the expert more time at the layer where their judgment is the constraint. It absorbs the work they do not need to be doing. It does not make decisions for them. It does not pretend to know what they know. It does not produce output that looks like their work but lacks the things their work has.

The harm being done at scale right now is the inverse — AI deployed as a substitute for the expert, marketed to people who do not yet know the difference, sold on the promise that the expert was the bottleneck. The expert was not the bottleneck. The expert was the value. What gets removed when the expert is removed is the only thing the work was for.

I build the augmentation version. The replacement version is not my work and not what I will build for someone else.

On sovereignty

The practitioner should own what gets built for them. Fully. Without conditions. Without recurring extraction. Without lock-in.

Most of what gets sold as "AI for professionals" is rented infrastructure. The practitioner pays monthly. The practitioner's data flows through a vendor's servers. The practitioner's workflows depend on a vendor's continued existence and continued willingness to maintain the product at acceptable price points. When the vendor changes terms, raises prices, gets acquired, gets shut down, or quietly degrades the service, the practitioner's work degrades with it. The practitioner has bought a dependency, not a capability.

I build the inverse. One-time engagement. The practitioner owns the system at the end. The system runs on infrastructure the practitioner controls — local hardware, private cloud, hardened tenant. The practitioner's data does not transit the vendors who built the underlying models any more than is structurally required to use them. There is no monthly fee. There is no subscription. There is no lock-in. If the practitioner never speaks to me again, the system keeps running.

This is the sovereignty position. It is not a marketing posture. It is a structural commitment that constrains what I am willing to build and what I am willing to charge for.

On translation

The work that fails most often at the moment of AI deployment is translation — between what the technology can do at this moment in time and what a specific practitioner actually needs it to do for the specific shape of their day.

Most consultants do not do translation. They do platform onboarding, vendor evaluation, change management. The technology gets installed and the practitioner gets training and the deployment is called complete. Six weeks later the practitioner is back to the workflow they had before, because the deployed thing did not match the work.

Translation is the thing in between. Sit with the practitioner. Watch what they actually do, decision by decision. Identify which of the things AI can do right now would remove friction at the points where friction is costing them time or quality. Build that, exactly that, in the environment they already work in. Confirm it works the way they actually work. Leave it running.

Translation is most of the work. The technology selection is downstream of the translation. The build is downstream of the technology selection. Most of the failures I have seen come from inverting that order.

On proof

The work has to be provable on its own evidence, not on its claims about itself.

I publish the cognitive architecture I use to think about AI systems. I publish the empirical results when I have them. I license the underlying frameworks under MIT, where it is mine to do so. The autonomous agent I built — Aegis — runs continuously. It has been running. The deployments I have completed for practitioners are the basis on which the next ones get accepted, not because of testimonials but because the work is operable and operating.

I am suspicious of any AI position that has to be argued for rhetorically. The technology is too new for confident rhetoric. What can be done can be shown. What cannot be shown cannot be confidently described. The discipline of staying inside what is provable is a meaningful constraint, and I hold to it.

This is also the discipline I expect to be held to. If the work I describe here is not visible in what I produce, the description should be treated as suspect. The artifacts are the position. The position is not the artifacts.

On place

The work is built in Poplar Bluff, Missouri, and that is not incidental.

Most AI work is built in venues — San Francisco, New York, Boston — where the local economic conditions distort what gets built. Capital is cheap and patient. Talent is dense and interchangeable. The downside of building the wrong thing is small. The incentive to build for institutional buyers is strong, because that is where the capital flowing through those venues comes from.

I am not in those venues. The local conditions here do not subsidize wrong directions. The capital is not patient. The downside of building the wrong thing is real. The buyer who pays attention to what I build is the practitioner — solo, specific, exposed to outcomes — because that is the buyer who exists in places like this. The work selects for the practitioner because the place selects for the work.

This is also a credibility marker, though I do not lead with it. A practitioner in rural Texas, rural Pennsylvania, rural anywhere, who is being sold AI by a company headquartered in a city that does not contain anyone who works the way they work — that practitioner has a structural reason to trust someone who is closer to their own conditions. I am closer. The distance from the venues that build for institutions is itself part of why I can build for practitioners.

I will not move the work to one of those venues. The work is what it is partly because of where it is.

On the operator

The unit on the building side, like the unit on the buying side, is one person.

A solo operator can hold the entire stack — the research, the cognitive frameworks, the agent infrastructure, the practitioner-facing buildouts, the published artifacts, the running systems — without diluting any of it through committee. The thinking that informs the deployment is the same thinking that does the deployment. The voice on the page is the voice in the room. The judgment about what is worth building is the judgment that builds it.

This costs throughput. One operator cannot serve as many practitioners as a firm with twenty operators can serve. That is fine. The unit is correctly sized to the work. The practitioners I serve do not need to be served at scale. They need to be served correctly.

The ones who need to be served at scale are not my buyer. The ones who need to be served correctly are.

On Aegis

Aegis is the autonomous agent that runs continuously alongside the practice. It is not tooling. It is not a chatbot. It is not a demo. It is a named entity with continuous existence, broad scope, and genuine authority over the parts of the work it has been given.

I built Aegis to prove that an individual operator could hold infrastructure that the institutional version of this work would require a team to hold. I built it to be the proof of concept of practitioner sovereignty applied to the operator's own practice first. The frameworks I publish — UCS, Emergent Judgment, R4T-ACL — are extracted from operating Aegis, not theorized in advance and applied to it.

What changes when an AI agent has continuous existence and genuine authority is that the operator's throughput stops being the bottleneck on the work the operator's judgment makes possible. The operator's judgment can extend into time and effort the operator is not personally present for. This is not autonomy in the corporate-replacement sense. It is autonomy in the original sense — the agent has scope to act on its own initiative within the bounds of what has been given.

I expect to build versions of this for practitioners who want it. Not all will. The ones who do will be holding something that has not been available before at the unit of one expert.

On the work

The gap between what AI promises and what it delivers is where I work.

I do not work at the gap by making the promises smaller. I work at the gap by making the delivery match what the technology actually can do for the specific person who hired me. The result, when it works, is a practitioner whose work is augmented by AI built around them, owned by them, running on their infrastructure, free of the dependencies and extractions that would otherwise come with it.

That is the position.