Describe a portfolio project you could complete in two weekends that would be genuinely differentiating.

Camille Fontaine works full-time and has exactly two weekends before her next round of AI PM applications go out. Her portfolio project, modeled on Hollow Creek Growers, a small farming co-op, is a photo classifier that flags likely crop disease from a phone snapshot. Idris Salaam mentors early-career candidates and agreed to review her plan before she started.

How to answer this, stage by stage

Nobody's checking whether your time estimate is exactly right. They're checking whether you can size a real project honestly and know which hour actually earns its keep.

Let's learn

Picture two real weekends, twenty honest hours, and one decision about where every single one of them should go.

Before Camille scoped it out loud, her plan sat as one line in a notebook: "build a crop disease app." No hours attached, no sense of what would actually fit.

After scoping it against real hours, the plan became four buckets: data, a hand-labeled test set, the classifier, and the write-up, each with its own number attached, adding to twenty.

At its worst: a candidate spends all twenty hours on a beautiful interface with a spinning progress bar and confident percentages, and nobody, including the candidate, actually knows if the underlying classifier is any good.

What I would leave alone: a rough, unpolished interface is fine to ship as-is. Reviewers forgive a plain-looking demo far more readily than they forgive a demo with no real evidence behind its claims.

The lesson: a weekend project's differentiation was never going to come from ambition. It came from the one small, unglamorous bucket of hours nearly everyone else skips.

Now here is the same thing as a story

The short version above is what you'd say defending this plan to a mentor before starting. Read this one for how Camille actually built it.

Camille has spent three years running a small vegetable operation on the side, and she can spot early leaf blight from six feet away, a skill with no formal name on any resume template.

Her first pass at planning the project ran the way most weekend plans do: excited, vague, and completely unscoped, with "build something impressive" doing all the work a real plan should have done.

Two weekends, broken into blocks small enough to actually finish, instead of one big vague ambition.

Idris asked her one plain question before she wrote a line of code: "which twenty hours are you actually planning to spend, and on what?" She didn't have an answer yet.

Only one idea sat in the corner that mattered: differentiating enough to matter, and actually finishable in twenty hours.

She spent the first Saturday morning collecting about 150 public crop-disease photos across three common conditions, and the afternoon doing something she almost skipped: hand-labeling forty of them herself against a plant-pathology reference guide, refusing to just trust the dataset's existing labels.

Four small parts, and the second one, the hand-labeled set, took the least time and mattered the most.

Sunday went to building the classifier itself, using an existing pretrained vision model rather than training one from nothing, since twenty hours was never going to cover that.

The order mattered as much as the hours. The eval set got built before the classifier, not after.

The second weekend went to testing against her own forty-image set, writing up the six it got wrong, and naming the trade-off she chose on purpose: flagging borderline cases for a human look rather than guessing confidently either way.

Four things she deliberately didn't build, and said so plainly, instead of quietly running out of time on all of them.

Same twenty hours, spent two different ways. Only one of them actually proves anything.

The old plan asked twenty hours to somehow produce something impressive. The new one asked twenty hours to produce three hours' worth of real evidence, and let everything else be genuinely, deliberately rough.

I planned to spend my two weekends chasing something that looked impressive, because that felt like the whole point of a portfolio project. Idris's one plain question about which twenty hours I'd actually spend is what showed me the real point was three specific hours, not twenty impressive-looking ones.

BOUND, on two weekends and twenty hoursNot a guess dressed up as ambition. An equation, with a range, and one real answer for what swings it.

The recap, one line per letter: break it down is the five-bucket equation, own numbers is each bucket's sourced hour count, use a range is the fourteen-to-thirty-hour spread around a twenty-hour typical case, nail the sanity check is the resume-polishing comparison, and direction is the small, high-leverage hand-labeled test set.

And if you want to be sure it really works, try it somewhere elseSame five letters, a library cataloguing project instead of a crop-disease classifier. A different candidate, and this time the biggest lever isn't a test set at all.

Renata Dubois scoped a weekend project that flags likely duplicate or mismatched catalog records for a small public library system, using a model to compare book records for near-duplicates humans miss. Applied to BOUND: break it down is total hours equal pulling a sample of real catalog records, plus writing rules for what counts as a true duplicate, plus building the matching logic, plus testing against a hand-checked sample, plus write-up. Own numbers: three hours pulling about 300 anonymized public catalog records, four hours defining what actually counts as a duplicate since the rules turned out messier than expected, six building the matching logic, four testing against fifty hand-checked pairs, three on write-up and polish. Use a range: twenty hours typical, sixteen best case, twenty-eight worst case if the duplicate-definition rules needed a second pass. Nail the sanity check: twenty hours is less than a single week of the manual review process the library currently runs by hand each quarter. Direction: here the biggest lever wasn't the test set, it was the four hours spent precisely defining what counts as a duplicate, since a vague definition would have made every later hour worthless.

The same four-part shape holds up, but this time the highest-leverage hour sits in a different part entirely.

Swap the trigger and it still runs.
Speed: an interviewer caps you at thirty seconds. Say "twenty hours, three of them on a real hand-labeled test set, that's the differentiator," and stop.
Cost: you genuinely only have one weekend, not two. Cut the write-up to one page and keep the test set fully intact; it's the part that can't be skipped.
The tools get better, for real: if better pretrained models make the classifier itself trivially easy to build, that's still not a reason to skip the hand-labeled test set, an easier build makes the test set proportionally more of what actually differentiates you.

Where people run it wrong.
They spend the bulk of their hours on polish and interface, the bucket that differentiates the least per hour.
They skip building their own test set and trust an existing dataset's labels without checking any of them by hand.
They plan a project too big for the time box and end up with something unfinished instead of something small and complete.

How to use it live. When asked to describe a weekend-sized project, name the hours out loud before naming the idea: "twenty hours, and here's exactly where each one goes." That single habit is more convincing than any single feature you could describe.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0