No client work appears on this site. The simulations and scripts ran on my own equipment, and the AI models ran on hosted services.
Some of it runs on the professional simulators I license. Some of it you can run on your laptop in under a minute. Both matter for different reasons: the first is what real work looks like, the second lets you check me yourself.
Two hours of a September morning in SIMetrix/SIMPLIS, cut to two minutes forty-one. The simulator on the left monitor, the AI on the right. The failed decks and the restarts are left in, and it ends on a number that was predicted before it was measured. Where the recording is compressed, the taskbar clock stays in frame, so you can see it.
Recorded 12 September 2026, 09:08 to 11:06. The measurement it ends on is on the compensator rule page. Music: Kimiko Ishizaka, Open Goldberg Variations, public domain.
A four-phase, 1 MHz, 119 A regulator held to Intel's VRD 10.1 guide, designed with nobody at the keyboard and released here as a datasheet, a design review and the decks that ran. Read in simulation. Nothing was built, and nothing was measured on a board.
Four phases, 1 MHz, 119 A on a 1.00 mOhm load line, held to Intel's VRD 10.1 in simulation. One page.
DATASHEET_4PH_1MHZ_2026-10-04.pdf, 248 kB, sha256 8a5b1b87
Sixty-three slides. The verdict, go to a first board with open calls, twelve findings, and the controller section that shows where a real part falls short and what a better one would buy.
DESIGN_REVIEW_4PH_1MHZ_2026-10-04.pptx, 2.4 MB, sha256 fdd06046
206 files. Nineteen SIMetrix/SIMPLIS decks with the prediction sealed for each, their checkers, schematics, netlists and receipts, nineteen figures, every document, a README and a manifest of hashes.
DOWNLOAD_4PH_1MHZ_DESIGN_2026-10-05.zip, 8.1 MB, sha256 0a344d5e
A different kind of evidence from the designs above. Thirteen models screened on analog and mixed-signal material by the same discipline, predictions hashed before the run and mechanical scoring, with the finding that the sandbox mattered more than the model: asked to compute a number, the ten open-weight models were cut off before answering in all 90 trials, and asked to write a script the harness runs, they got 81 of 90 right.
Every number on this site carries one, and a tag is never upgraded because a result looks right. A check that cannot decide returns NOT EVALUABLE, never a pass.
The worked example. A converter that fails its linearity spec and passes after calibration; a yield that goes from 2 % to 97 % with the trim depth; and a routine that produced a 0 % yield when one check was removed, completing without an error either way.
MEASURED runs on your laptop
A published compensator rule, twenty years old, reproduced from its own figures and found incomplete for the multiphase case. The gap was derived before the source text was read, then confirmed at two, four and eight phases against one simulated and one measured arbiter, with the alternative explanation written down first, tested, and refuted.
SIMULATED on the simulator I license
119 A, 1 MHz a phase, held to Intel's design guide line by line at typical part values, with the misses at datasheet limits listed on the case page. The engineer set the target and handed over the design calls on purpose; 22 of 31 recorded decisions were the AI's, every one on record, reviewed cold five times by a second model. Read in simulation, nothing built. Forty-one wrong turns on record.
SIMULATED on the simulator I license, files on request
Every item below is a real defect from building the work on this site. Each one produced output that looked entirely reasonable. Each one is now caught by a specific, mechanical check, and those checks are the actual deliverable of an engagement. They are what your team is taught to write in week four, and they are written down and yours to keep.
Notice what is not in that list: a bigger model, a better prompt, or a smarter tool. Every one of these was caught by something small, mechanical and boring. A known-answer probe, a second configuration, a separated metric, a check on the checker.
The automation makes it affordable to run checks at a scale nobody would do by hand; the checks are what make the output trustworthy; and an engineer still decides what the results mean.
The checks are the deliverable. Week four of an engagement is where your team learns to write them against your own flow, and the discipline, the standing rules, the gates, the claim-tagging practice, is written down and stays with you whether or not anything else does.