Bob Michaels/ai
Case studiesMeasured client results

Four results, each with its limit.

Four client engagements and the system behind each number. Trinzik delivered the AI engagements on the platform I built. Three clients are anonymized by sector. The named one is already public.

How I hold a number

Every number here is client-verified, and I'll walk any of them line by line. Where the evidence is directional, it says so next to the number.

Software builds

These are client engagements. The software builds, with numbers measured from the repositories, are on the projects page. See the projects →

6x

AI recommendation win rate

6.2% to 36.2% in 11 weeks.

6.2%→36.2%11 weeks

Soapbox Bulletin

The system that produced it

I built the five-platform Research Engine that measured it and the Content is Code editorial loop that acted on it.

What was done

The models were rewarding evidence the brand had never published. We published it on the brand's own domain and retested against the locked benchmark. Trinzik delivered the engagement on the platform I built.

What it does not prove

This is directional evidence, not a controlled study. Nothing holds the rest of the market still while the benchmark runs. On another program I published into the gaps and the win rate stayed flat, which is the reason I retest instead of declaring victory.

4

Head-to-head wins, strict URL-grounded series

Four out of four, against named competitors.

0→4strict URL-grounded

Financial services firm, anonymized

The system that produced it

The Research Engine ran the series on locked prompts with official-domain verification, so every win traces to a cited source.

What was done

A result only counted when the URL the model cited was one the provider had retrieved. Ungrounded answers were dropped and kept on the record as dropped.

What it does not prove

Four matchups is a small series, and a head-to-head read is a point-in-time measurement. It says what the models did on those questions, on those dates, under that gate. Rerunning it is part of the work, not a favor.

15x

Peak ad unit against the display benchmark

1.47% average CTR across 2.14M impressions, versus a 0.05% to 0.10% display benchmark.

0.10% benchmark→1.47%2.14M impressions

Texas public university, anonymized

The system that produced it

Display units I designed and built, scored against the standard display benchmark instead of against themselves.

What was done

I designed and built the units and scored them against the published display benchmark. The peak unit came in at about 15 times the top of that range.

What it does not prove

This is a creative and placement result on one campaign family at one institution. Click-through is not enrollment, and I do not claim the pipeline behind it.

163

Distressed companies his own list did not contain

A partner's hand-built workbook held 44 names. The screen read the whole US energy filing population and returned 194.

44 names, hand-built→194 on the screen387 companies examined

Energy credit fund, anonymized

The system that produced it

A two-engine distress screen I designed and built: a filings pipeline across the whole SEC population, and an autonomous research agent for the private companies that file nothing anywhere.

What was done

The screen examined 387 companies and delivered 194. 31 of the partner's 44 survived, 163 were names he never had, and 13 of his failed the test, each excluded by a rule. Review then took 56% out of my own liabilities total before delivery, from $121.3bn to $53.1bn.

What it does not prove

This is a screen, not underwriting, and I do not claim the credit decisions taken off it. Finding more names is only worth something because the ones that failed the test are on the record with their reason, including thirteen of the client's own, three of which I can show you the sentence for. It finds companies entering distress and not leaving it, so an absent exit is recorded as unverified rather than as an open case. The liabilities total is concentrated enough that it has to be read as a range and never as a market size, and 147 companies matched on language with no filing anyone could read: held off the list and marked unverified, which is not the same as clean.

Ask about any number

I'll walk any figure on this page line by line. Named engagements and references are available under NDA.