How do you build a credible AI PM portfolio with no AI job experience?

Mattias Berg spent eight years as a veterinary technician, the person who first looks at every animal that walks through the door. His portfolio project is an intake-triage tool modeled on Bramblewood Animal Hospital: it flags which incoming cases likely need to be seen before their scheduled slot. Priyanka Shah hires AI PMs at a health-tech company, and Mattias applied to nine of her team's postings before landing one callback.

How to answer this, stage by stage

This isn't a motivational question about persistence. It's asking whether you can diagnose exactly where your applications are actually failing, instead of assuming you already know.

Let's learn

Picture a career switcher who's genuinely good at the underlying judgment an AI PM role needs, with a resume that says nothing about it.

Before Mattias built anything, his resume read like every other vet tech's: patient care, client communication, inventory management. Nothing about it signaled he could reason about a model's failure modes, because nothing in his actual job title ever touched one.

After he ran the diagnosis, the real picture came apart into two very different stories: five rejections that turned around the moment he added one built artifact, and two that stayed rejections no matter what, because those specific roles genuinely required production ML experience he didn't have yet.

At its worst: a genuinely capable candidate spends six more months believing the entire problem is a missing job title, when the actual, fixable gap was one weekend's worth of building.

What I would leave alone: the two rejections from roles genuinely requiring years of production ML deployment were correct rejections, not a portfolio failure. No artifact closes a real experience gap for a role that specifically needs it.

The lesson: "I don't have the experience" and "I haven't proven the judgment yet" feel identical from the inside, but only one of them is actually fixable in a weekend.

Now here is the same thing as a story

The short version above is what you'd say diagnosing your own stalled job search. Read this one for how Mattias actually found the real cause.

Mattias could glance at a dog limping into the clinic and know, before a vet even looked, whether it needed to be seen in the next ten minutes or could safely wait.

His first round of nine applications went out over six weeks, each with the same resume, each describing his real experience in the only language he'd ever used for it.

Almost every rejection happened at the very first gate. Nothing about his actual judgment was ever tested at all.

Nothing dramatic marked the turning point. Just a slow accumulation of silence, application after application, until a friend who'd made the switch a year earlier asked him one question: "when did you last hear back from any of them?"

The funnel narrowed almost entirely at the second box. Everything after it was nearly empty.

He spent a weekend on the smallest real version of an idea he'd had for years: a tool that flags incoming cases by likely urgency, using the same signals he used in his head every day.

Four small parts, built in a weekend, doing more work than six weeks of unanswered applications.

He sent the exact same resume, with one link added, to a fresh batch of eight similar roles, deliberately as close to a controlled test as a job search allows.

Three suspects, and the test wouldn't finish until one of them was actually confirmed.

Five replies came back within two weeks, more than the previous six weeks had produced combined, and two of the remaining three rejections named production ML experience directly, a mismatch no artifact was ever going to fix.

The evidence pointed at one branch clearly. The other two suspects turned out not to be the real cause at all.

The same portfolio worked completely differently depending on which kind of team was reading it.

The old search treated every silence as proof that his background was the problem. The new one separated a fixable gap in proof from a real, unfixable role mismatch, and stopped wasting effort on the second one.

I spent six weeks assuming the missing job title explained everything, because it was the most obvious difference between me and other candidates. Watching one weekend's build change five out of nine outcomes is what showed me the real gap was never the title. It was that nobody had anything of mine to actually check.

TRACE, on a job search that isn't convertingNot a confidence problem. A diagnosis, with three suspects and one real test.

The recap, one line per letter: timeline is the drop happening at the resume screen, recut is startup versus big-tech teams filtering differently, assume nothing is ruling out formatting first, cause candidates is the three named suspects, and evidence test is the same resume with one artifact added, run as a real comparison.

And if you want to be sure it really works, try it somewhere elseSame five letters, a port operations career switch instead of veterinary work. A different candidate, and this time the real cause is something else entirely.

Jonas Eriksson spent six years as a container-yard supervisor before trying to move into AI product work, building a portfolio project around a berth-scheduling tool for a regional port. Applied to TRACE: timeline showed his rejections weren't concentrated at the resume screen at all, several reached a phone screen and stalled there instead. Recut: the drop was consistent across company size, which ruled out a segment-specific filter. Assume nothing: his resume was reaching humans fine, confirmed by two friends who forwarded it internally. Cause candidates: no AI vocabulary, no judgment artifact, or a communication gap in the phone screen itself. Evidence test: a mock phone screen with a former recruiter revealed the real cause, Jonas answered every technical question well but never once named a trade-off he'd have made differently, which read as reciting facts rather than showing judgment.

Same three suspects, different confirmed cause. Jonas already had an artifact. What he lacked was practice naming the judgment behind it.

Swap the trigger and it still runs.
Speed: an interviewer wants your diagnosis in one sentence. Say "check where the drop happens before assuming why," and stop.
Cost: you can't afford weeks of testing different resume versions. Send just two versions, one with an artifact and one without, to a small, similar batch, and compare.
The market gets better, for real: even when AI roles get less competitive overall, the same diagnostic habit still matters, more openings doesn't tell you which specific gap is costing you callbacks.

Where people run it wrong.
They treat every rejection as proof of the same cause, without checking whether the drop happens at the same stage each time.
They keep applying to roles that genuinely require experience they don't have yet, mistaking a role-fit mismatch for a portfolio problem.
They test everything at once, new resume format and new artifact and new roles simultaneously, so nothing about the result is actually attributable to one change.

How to use it live. When your own job search stalls, ask exactly one question before doing anything else: at which stage does it stall, every time. That single answer tells you whether to fix your resume, build an artifact, or change which roles you're aiming at.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0