How would you use a portfolio project to open a conversation with a hiring manager?

Larkspur Realty Group is a fictional real estate brokerage. Nikolai Bauer built Framecheck there: it scores a listing's photos for lighting, clutter, and framing, then tells an agent which one to feature first. Isolde Van Praet, a hiring manager, is the first person Nikolai practices his opening line on.

How to answer this, stage by stage

Nobody is grading whether your opening line sounds impressive. They're grading whether it survives the very first question back.

Let's learn

A listing-photo scorer is a tool that looks at a set of real estate photos and tells an agent which one is actually going to make someone stop scrolling.

Nikolai had three numbers he could open a conversation with. First: "our model hits 91% top-3 accuracy picking the best photo." Second: "listings using our top-scored photo got 31% more clicks in a live test." Third: "cut an agent's photo-picking time from 12 minutes to under 2."

Only one of those three numbers came from an independent source. The other two lived entirely inside Nikolai's own head and his own stopwatch.

Here's the turn: the extra impressiveness of the 91% figure was never the point. What mattered was whether the number could survive being poked at, and the flashiest claim was the one most likely to fall apart under one direct question.

At its worst: Nikolai opens with the 91% figure, an interviewer asks one plain question about his eval set, and the rest of the conversation becomes about defending a shaky number instead of discussing the actual product decisions behind Framecheck.

What I would leave alone: the 91% and time-saved numbers don't need to be deleted. They're fine to mention later, clearly labeled as self-tested, once the conversation has real trust behind it.

The lesson: the strongest opening line isn't the one that sounds best. It's the one that's already been tested against the exact question that's coming next.

Now here is the same thing as a story

The short version above is what you'd say explaining your opening line out loud. Read this one for how Nikolai actually picked his.

Nikolai Bauer spent three years staging homes for open houses before he ever trained a model, and he can tell within a glance which photo of a listing is going to stop someone mid-scroll.

He'd practiced his opening line a dozen times before his first real interview, always reaching for the 91% accuracy figure, since it was the biggest number he had.

Two numbers, both true. Only one of them was ever checked by someone other than Nikolai.

A friend doing a mock interview asked the obvious question: "How'd you get to 91%?" Nikolai explained he'd hand-picked 50 photos from listings he liked, run the model, and counted how often it agreed with his own judgment.

The exact same number reads two completely different ways, depending entirely on who chose the 50 photos.

His friend didn't say it was wrong. She just said, plainly, "that's your opinion of your own model, dressed up as a statistic." Nikolai sat with that for a while.

Checking the eval set behind each claim is the step that actually separated a real number from a convenient one.

Going back through his own materials, he found the click-through number he'd almost forgotten about: Larkspur's marketing team had run a real, live test across six weeks of published listings, comparing click rates on listings that used Framecheck's top pick versus ones that didn't. Thirty-one percent more clicks, tracked entirely outside his own hands.

All four of these were true about the click-through number. None of them were true about the 91% figure.

Rebuilding his opening line around the click-through result took Nikolai about twenty minutes, mostly digging up the original marketing report to make sure he could still cite it accurately. Practicing it against the same hard question took another mock interview to get comfortable.

The click-through claim was the rare one that landed in the corner that actually matters: impressive, and solid.

Leading with the 91% number had felt natural, since it was the biggest, roundest figure Nikolai had. It stopped feeling like the right choice the moment one honest friend asked where it actually came from.

I opened with my biggest number because it was the most impressive one I had. It took one plain question from a friend to see that impressive and defensible were never the same test.

AUDIT, turned on your own opening lineNot a critique of someone else's report. AUDIT is what tells you which of your own claims is safe to lead with.

Run every candidate opening line through these five questions before you ever say one out loud.

The recap, one line per letter: ask who verified it separates self-graded from independently checked, uncover the eval set separates a curated sample from a real population, demand the version pin tests whether a number can be traced back, isolate what's missing checks for a stated limit, and test it yourself is the final, hardest check.

And if you want to be sure it really works, try it somewhere elseSame five letters, a fisheries co-op instead of a real estate brokerage. A different building, and this time the flashiest number is the one an interviewer already suspects.

Brackwater Fisheries Collective is a fictional fishing co-op. Halvard Espen built an AI tool there that predicts the best days to bring boats out based on weather and catch history. Junko Arai reviews his opening claim.

Mapped onto AUDIT: Halvard had two candidate openers, "our model beat the co-op's own forecast by 22% on trip yield" and "fishers using our recommendation saved an average of 40 minutes of planning time each morning." Asking who verified each: the yield number came from the co-op's own catch-logging system, tracked independently of Halvard. The time-saved number came only from three fishers he personally interviewed. Uncovering the eval set: the yield comparison ran across an entire fishing season's real trips, while the time-saved figure came from three conversations. Demanding the version pin: the yield number was tied to a specific model version locked before the season started, fully traceable. Isolating what's missing: the time-saved claim had no stated sample size in Halvard's notes until he went looking for it. Testing it himself: he could pull the yield comparison straight from the co-op's own records, live, in front of anyone. He led with the yield number, and kept the time-saved story as a supporting anecdote instead.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "here's the one number I could pull up and defend live, right now," and open with that.
Cost: there's no time to build a full independent test before tomorrow's interview. Say plainly which numbers are self-tested and which aren't, rather than let one blur into the other.
The model gets better, for real: if Framecheck's accuracy improves later, that's still worth mentioning, but it doesn't change which claim you open with, since the audit is about verification, not about how good the number is.

Where people run it wrong.
They lead with the biggest number instead of the most defensible one, and get stopped by the very first follow-up question.
They assume any claim they can state confidently is safe to open with, without ever checking who actually verified it.
They delete the shakier claims entirely instead of keeping them in reserve, clearly labeled, for later in the conversation.

How to use it live. When someone asks how to open a conversation with a hiring manager, ask yourself one question first: which of my claims could I reproduce right now, on the spot, if asked. Open with that one.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0