We find out whether your AI actually works.
Including ours.

7 → 2

Our own rigor cleared five of the seven people we had named.

An earlier version of a workforce analysis put seven names on a list. Rebuilding it with proximity calibration withdrew five of them, including the strongest lead, who turned out to have been standing at a customer's front door. The two that remained were left with the client to review internally. We did not, and do not, characterise anyone on that list as having done anything wrong. Every bug we fixed made the findings smaller.

31% → 43%

We restated our own productivity finding, upward.

Our first pass told a client their field technicians were 31% productive. Our own follow-up audit found that 57% of the supposedly unproductive hours were within 100 metres of a real customer site. We corrected the figure to 43% and withdrew five of seven names from a suspected-moonlighting list.

3 models

We built three ML models and refused to ship any of them.

Measured properly, the accuracy turned out to be an artifact of leaked labels and circular targets. We flagged all three as untrusted and kept them shadow-only rather than put a confident-but-wrong model in front of a real decision.

Don't build it

We formally advised a client not to build AI at all.

Not yet, and not on that data. We put it in writing: finish the data-trust work first, because a model trained on numbers nobody can defend produces confident answers nobody should act on.

$1,500 → $234

We told a client our cost estimate was more than four times too high, unprompted.

We estimated a client's monthly AI token cost at $1,100 to $1,500. Actual spend came in at $234. We told them.

~30 fixes

We found a deployment that had silently reverted itself.

The build running in production was an older snapshot. It had quietly undone roughly 30 correctness fixes and dropped nine named data feeds, which had been returning nothing for weeks while the health badges stayed green. We recovered them and added a staleness watchdog so a feed going quiet raises an alarm instead of a green light.

906 of 1,290

We graded our own data by how it was obtained.

In a product we built, we added a confidence spine that grades every listing by its provenance, scraped, inherited, or model-derived, and serves it by trust tier instead of silently hiding the weaker rows. Separately, that work established how many active listings were genuinely bookable, meaning a real future session and a real price: 906 of 1,290.

Sent ≠ sent

We found a Send button that lied.

An approval console marked time-sensitive notification campaigns as sent without ever calling the mail vendor. Nothing went out, and the audit trail said everything had. We fixed the send path and added an expiry, so a stale campaign now dies quietly instead of reaching someone weeks after the event it describes has already ended.

2 disproved

We published the experiments that disproved us.

A provisional patent application covering Edwin, the AI memory product our founder is building (in pre-release), was filed with 10 claims validated against 35 experiments. Two of those experiments disproved what we expected. The filing says so.