Describe how to present a failed project in a portfolio.

Thornfield Realty Partners is a fictional commercial real estate firm. Mireille Kanu built Clearlease there as her portfolio project: it reads scanned lease PDFs and pulls out ten key fields, rent terms, renewal dates, termination clauses, for a firm's analysts. The pilot, five analysts, four weeks, ended quietly. Walt Bregman, a hiring manager, is the first person she's shown the real write-up to.

How to answer this, stage by stage

Nobody is grading whether your project worked. They're grading whether you can say, precisely, why it didn't.

Let's learn

A lease-abstraction tool is a product that reads a scanned contract and pulls out the handful of facts someone would otherwise hunt for by eye: rent, dates, exit terms.

Clearlease launched to five analysts at Thornfield Realty Partners. In week one, all five opened it every day. Extraction accuracy across all ten fields it pulled sat at a strong 94%.

By week four, one analyst out of five was still opening it. The dashboard showed a clean 94% the whole time, unchanged, sitting right next to a usage number that had quietly fallen off a cliff.

Here's the turn: the extra mistakes hidden inside that 94% were never the real problem. Analysts didn't quietly distrust the whole tool. They trusted nine of its ten fields completely, and manually redid the tenth, every single time, without ever saying a word about it.

At its worst: Mireille nearly left Clearlease off her portfolio entirely, since "the pilot failed" felt like the whole story, and she assumed a failed pilot had nothing worth showing.

What I would leave alone: the underlying extraction model and its OCR pipeline were genuinely fine and didn't need retraining. The fix was entirely about surfacing a confidence number the model already had, not making the model smarter.

The lesson: a failed project isn't a project with nothing to show. It's a project where the real cause is still sitting there, uncut, waiting for someone to actually look.

Now here is the same thing as a story

The short version above is what you'd say walking a hiring manager through this project cold. Read this one for how Mireille actually found the real cause.

Mireille Kanu spent three years as a commercial leasing analyst before she ever wrote a line of code, and she could spot a bad renewal clause on a scanned lease faster than most of her old team could open the file.

Clearlease's first week went well. Analysts opened it every morning, pulled ten fields per lease in seconds instead of the twenty minutes it used to take by hand, and told her it felt like magic.

The UI cleanup in week two looked like an improvement on paper. It's also exactly when the real trust broke, two weeks before anyone noticed the usage number moving.

Then the drop-off began, in three quiet beats. Week two, one analyst stopped opening it daily and went back to checking leases by hand "just for this batch." Week three, two more followed, each with their own private reason that sounded reasonable on its own. Week four, only one analyst was still opening it at all.

The trigger wasn't a single bad output or an angry email. It built up slowly enough that Mireille almost missed it, the way these things usually do.

Both numbers were true at the same time. Only one of them was the one that mattered to the people using the tool.

Before she blamed anyone, Mireille checked whether the usage logs themselves were even trustworthy. They were. The drop-off was real, not a tracking artifact.

Three real suspects, and only one of them survived a real check.

She pulled the click logs next, expecting to find analysts double-checking every field out of general suspicion. Instead, she found something narrower: every analyst, every time, manually retyped the renewal date and left every other field untouched.

Renewal date sat at position nine of ten, sorted alphabetically, with no confidence flag telling anyone it was the weakest field in the list.

The old design asked every field to compete equally for an analyst's trust, sorted by nothing more meaningful than the alphabet. The new one, which she sketched but never got to ship before the pilot ended, put renewal date first and flagged it whenever its own confidence dropped below its known accuracy.

Renewal date sat in the one dangerous corner: low accuracy, and nothing on screen telling anyone not to just assume it.

Sorting the fields alphabetically had felt like a harmless, even tidy choice back when there were only three fields to show. It stopped being harmless the day the list grew to ten and the one field analysts actually leaned on landed near the bottom, indistinguishable from nine fields they could trust completely.

I sorted those ten fields alphabetically because it was simple and I never revisited it once the list grew. It took watching four out of five analysts quietly stop opening the tool to see that "simple" and "safe" had stopped being the same decision.

TRACE, the real diagnosisNot a eulogy for a failed pilot. TRACE is what turns "it didn't land" into a cause someone can actually check.

The click-log check is the step that actually separated a vague story from a real cause.

The recap, one line per letter: timeline is the break landing two weeks before the number moved, recut is the 94-versus-71 split, assume nothing is ruling out a logging bug first, cause candidates is the three named suspects, and evidence test is the click logs that convicted one of them.

And if you want to be sure it really works, try it somewhere elseSame five letters, a school district's attendance tool instead of a lease. A different building, and this time the real cause is a calendar, not a field.

Windmere Veterinary Group is a fictional veterinary chain. Emory Castellanos built an AI dictation assistant there, one that turns a vet's spoken exam notes into a structured chart entry. Priya Osgood reviews his write-up of the tool's rocky pilot.

Mapped onto TRACE: the timeline shows usage held steady for three weeks, then broke sharply in the fourth, right when the clinic's slow season ended and appointment volume doubled. The recut splits accuracy by appointment type instead of by field: routine checkups transcribed at 96%, while urgent same-day visits, spoken faster and with more interruptions, transcribed at 68%. Assume nothing rules out a software update as the cause, since the dictation model itself never changed that month. The cause candidates are three: vets rushing their speech under higher volume, more background noise in a busier clinic, or the model genuinely struggling with urgent-visit vocabulary specifically. The evidence test compares transcription accuracy against appointment length rather than volume, and finds accuracy tracks tightly with how many words per minute a vet spoke, not with how many patients were in the waiting room.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "94% overall, 71% on the one field that mattered, and click logs proved that's exactly the field people stopped trusting," and stop.
Cost: there's no budget to rebuild the field ordering before the next pilot. Say so honestly, and start with just adding the confidence flag to the existing layout, since even that alone would have caught the problem.
The model gets better, for real: if Clearlease's renewal-date accuracy improves to 90% next quarter, that's still worth mentioning, since it proves the diagnosis was right and the fix actually worked.

Where people run it wrong.
They summarize a failure as "it didn't get traction," which tells a reader nothing they could check.
They blame the users first, before ever ruling out a measurement or logging problem.
They name one plausible cause and stop, instead of naming several and running the one test that actually separates them.

How to use it live. When someone asks you to present a failed project, ask yourself one question first: what's the one check that proved your real cause, and not just your best guess. Lead with that check, not with an apology.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0