The labs can't staff it
The AI labs are hiring forward-deployed engineers faster than they can find them. The work isn't the model. It's getting the model to behave inside one company's real process.
Off-the-shelf AI works where the rules are light and the systems are simple. Banks, hospitals, plants, insurers and agencies are none of those. The answer is already in your paperwork; it is stuck behind old systems, strict rules and decisions that can't be wrong.
We find the one step everything waits on, build the smallest thing that clears it on the systems you already run, and show you the hours.
Every automated step has to be explainable to an examiner, an auditor or a compliance team.
Core platforms that predate modern APIs, and a different setup at every site, branch or state.
Money, patients, claims and public funds. The machine prepares; a person decides.
The five where it is needed most
Also built and shipped
The AI labs are hiring forward-deployed engineers faster than they can find them. The work isn't the model. It's getting the model to behave inside one company's real process.
"Process claims twice as fast" isn't a buildable job. Three days lost because a signature page was missing on Tuesday and nobody noticed until Friday is. We work at that level.
We pick the step where a small build moves the most work with the least authority handed to the machine. Credit, coverage, clinical and legal calls stay with your people.
That's enough to tell you, in writing, which step is costing you the most and what it would take to clear it. No build until the number is on paper.
The first five are where regulation, old systems and high stakes overlap, and where embedded engineering is needed most. Each page shows the bottleneck we look for first, the fix that usually clears it, and the number we put on the table before building.
Where it is needed most
Also built and shipped
A full engagement runs about thirty days. When the files are clean and the operator is available, the first two weeks compress to two days.
| Week | What happens | What you get |
|---|---|---|
| 1 | We pull the last 10–20 completed cases and reconstruct what actually happened. Fast cases against slow cases. Issues sorted into types with rough counts. | A list of your top problems with a rough cost on each. |
| 2 | We sit next to the person doing the work. The real process is never the one on the diagram. | A one-page claim: this step, these hours a month, these guardrails. You approve before anything is built. |
| 3 | We build the smallest fix that clears the step, giving the model only the data it needs. We test it against about 50 past cases with known answers, clean and ugly, and rerun after every change. | A working fix and a pass rate. |
| 4 | Two or three of your people use it while we watch. We fix the loop. | A measured result and someone who still answers the phone. |
The fix flags, drafts, sorts, and fills. A person approves anything that moves money or carries legal weight.
If the first fix saves 20 hours a month instead of 2,000, we go back and find out why. The target is the step everything else waits on.
Size and safety are visible on paper before week three. If the number isn't there, we don't build.
A fix that scores well on old files can still fail in production. We stay to see what real people do with it.
Centripetal is operated by Dave Motheral out of Fort Worth, Texas. The work below is run on the loop described here, and every number comes from the record.
June to August 2026, same step, three cycles. 605 statements and $1.98M in scope in August; 483 tie to the receiver's register to the cent. September regressed and is on the board with a reason on every held file.
A 220-item holdout set, attested by hand. Zero invented items on every run, checked by the database, not the builder. A confirm draw the engine never trained on scored 0.986. One sub-gate still open; it says so in the record.
The operating system ships its own code through the same loop: 1,439 dispatches across 38 codebases, 715 with a pull request. Seventeen stopped at a named gate: spending, access rules, outside sends, deletions, secrets. At a three-hour-per-change basis that is roughly 700 hours a month of engineering, estimated.
Each one comes from a table in the record, not a slide. The throughput line is the same step measured across cycles. The eval line is a holdout the builder could not see. The ownership line counts work that reached production with a review, and the small number that was stopped on purpose.
About 140 hours a month on the mineral step: 660 statements at a 15-minute manual basis, less the review time people still spend. About 280 hours once per legal matter: 583 documents at a 30-minute review basis. Estimates from volume and a stated basis, labelled as such. Where a lane has not run the loop, its page says so.
6,090 GitHub contributions in the last year (github.com/dmotheral3-eng), 107 repos, about 3,900 deployments (about 1,100 to production), 17 databases with 3,000+ tables, and AI agents and MCP servers in client work across 18 industries.
One operator covers many industries because the method is written down and partly enforced by software. An industry with no cases pulled, no step chosen, no test set and no pass rate has not been debottlenecked yet.
On the morning of 10 October 2026 the owner sent a twenty-minute video about the forward-deployed engineer job and said: this is what I do. What follows is the clock, read from the record, not from memory.
Eight in the first month. Six hundred and ninety-two in August 2026. The jump in June is the month the parse step was moved into the loop.
Each one is a ticket-sized change with a branch and, where it finished, a pull request. Median twelve minutes from fired to done.
A decision here is a row: what was chosen, why, by whom. The record keeps them so the next person does not re-decide.