← Case studies

Private Fleet Net Zero

Building an autonomous system that catches what a manual check misses.

Private Fleet Net Zero connects freight buyers with private fleet trucks that return empty on 80% of their trips.

Previously, testing the marketplace meant manually completing a freight transaction from both sides. Prozan AI built software that now simulates the entire matching process every morning, checks that information stays consistent, and reports issues directly to the CEO. Within its first week, it identified two bugs missed during manual testing.

Russ Jones, co-founder and CEO of Private Fleet Net Zero
“I was impressed by how quickly Noah understood our business and got the system running. Within two weeks, we had an automated check that was already catching things we’d missed manually. It’s taken a repetitive task off our plate and gives us a much better picture of what’s happening in our marketplace before the workday even starts.”
Russ JonesCo-founder and CEO, Private Fleet Net Zero

The problem

A 13-step check by hand, every morning

Every morning, someone at PFNZ had to test the marketplace like a real customer. They’d log in as a fleet, enter a truck’s route, log out, switch to a buyer account, and enter a shipment. Then they’d wait for a match, open the emails, review the offer, send a counteroffer, accept it, and confirm the deal.

Thirteen steps, two accounts, multiple emails and pages. All repeated by hand to make sure the marketplace was working.

And because PFNZ releases new software frequently, a check that passed yesterday didn’t guarantee everything would work today. A change to one screen could affect matching, emails, or the pages customers use to complete a shipment.

Even completing every step wasn’t enough to catch everything. A pickup date could appear correctly in an email but incorrectly on a customer-facing page. The transaction would still go through, leaving conflicting information that was easy to miss when checking each screen individually.

The deliverables

What lands before the workday

The morning email open in Gmail on an iPhone at 6:04 a.m.: subject 'Matched, but wrong pickup day (morning check, Nov 9)', from PFNZ Morning Check; PFNZ's logo, 'Two pages show the wrong pickup day', and every value by step. Test data.
The morning email. Green, yellow or red in the subject line. Inside: what happened in two sentences, then every value by step, the wrong ones in red, with screenshots. On the CEO’s phone. Test data.
The run log's History tab: every morning as a colored square on a calendar, green, yellow or red, with monthly totals. Sample data.
The run log. A Google Sheet the CEO already knows how to use. The History tab lays every morning on a calendar, so a bad release shows up as a streak, not a rumor. Shown with sample data.

Also delivered

A one-page guide
Explains the daily reports, how to review test results, and how to pause or adjust the automated check.
Commented code and a README
So PFNZ’s own engineers can follow every step, run it on a laptop or move it into PFNZ’s Google Cloud account.
Monthly support
Ongoing maintenance, email support with quick response times, and AI strategy consultations to explore new opportunities for PFNZ.

What it found

The kind of error a manual check can’t see

In its first week, the check flagged a pickup date that read Saturday in every email but Friday on two customer-facing pages. A person checking one screen wouldn’t see it, and a fleet planning from the wrong page shows up a day early. The CEO had a report with screenshots that same day and passed it to his team.

One test deal’s pickup date, in all 8 places it appears

Fleet
  • Offer emailSaturday
  • Offer pageFriday
  • Counter emailSaturday
  • Matched! emailSaturday
Buyer
  • Loads screenSaturday
  • Counter emailSaturday
  • Acceptance pageFriday
  • Matched! emailSaturday

The engagement

Agreement to first report in 14 days

  1. Oct 14

    Discovery

    The CEO walked through the morning check his team did by hand.

  2. Oct 23

    Agreement

    A written scope: what the check covers, what it never touches, what counts as done.

  3. Oct 28

    Build

    On PFNZ’s test accounts only, never real customers or fleets.

  4. Nov 6

    First report

    A full run, reported to the CEO.

  5. Nov 7 on

    Every morning

    On a schedule, with monthly support.

For engineers

Plain code decides. The model explains.

Whether PFNZ passed is a yes-or-no question with one right answer, so deterministic code makes that call. Gemini takes context from the run and writes the two or three plain-English sentences on a bad morning, and code checks those too: a note that names a number, day or month not in the run’s facts is thrown out.

How the PFNZ morning check is built Cloud Scheduler starts a Cloud Run job every morning. The job signs in to PFNZ's website as a test fleet and a test buyer in Google Chrome, reads the test inboxes through the Gmail API, compares every detail in plain code, has Gemini write a note only on yellow or red mornings, then writes PFNZ's Google Sheet, emails PFNZ's CEO and cleans up its test data. Credentials come from Secret Manager. New versions are built by Cloud Build and stored in Artifact Registry, tagged by commit. 6:00 a.m. alarm Cloud Scheduler every morning, with retries Locked vault Secret Manager Managed credentials CLOUD RUN JOB a fresh machine for each run 1 · Uses PFNZ like a person Python · Playwright · Google Chrome 2 · Compares every detail plain code · 8 places · 12+ details decides green, yellow or red 3 · Writes the note Gemini · Vertex AI yellow or red only · a guard checks its facts 4 · Reports, then cleans up Sheets API · Gmail API Cleans up test data PFNZ's website a real browser session Isolated test environment Two test inboxes Gmail API offer, counter, Matched! emails PFNZ's Google Sheet Today and History tabs Email to the CEO color in the subject, screenshots HOW A NEW VERSION SHIPS The code Python · 66 tests Cloud Build packages each version Artifact Registry labeled with its code change

Tested the way a customer sees it

A real browser and real inboxes, so a page that looks wrong to a fleet looks wrong to the check.

Protects live marketplace data

Every test record is tracked and removed after the check, ensuring automated testing doesn’t interfere with real customer activity or leave behind false marketplace data.

Keeps credentials secure

Passwords and API keys are stored securely outside the code, with each component given only the access it needs.

Built with: the stack, layer by layer

Starts it

  • Cloud Scheduler starts the run every morning, with retries
  • Cloud Run runs it on a fresh Google server

Runs the deal

  • Python the whole system, with 66 tests
  • Playwright + Chrome drives PFNZ as the fleet and the buyer
  • Gmail API reads both test inboxes, sends the report

Explains and reports

  • Gemini on Vertex AI writes the note on a bad morning; never decides pass or fail
  • Sheets API writes the run log

Keeps it safe

  • Secret Manager keeps credentials out of the code
  • Cloud Build + Artifact Registry package and label every version

What could work better in your business?

Have a process that takes too much time or keeps causing problems? Prozan AI builds and maintains custom systems to make it work better.

Book a free 30-minute call →