Private case study

Enter it below to read on. The rest of the portfolio is open from the menu above.

TalentReef · 2026

I designed TalentReef's first AI feature. It tells hiring managers what's worth knowing about an applicant, and admits when there's nothing there.

Role

Product Designer

Timeline

Feb–May 2026

Team

PM, design manager, engineering

Platform

Web · desktop and mobile

What I owned

NPS analysis · competitive analysis across 20+ products · concepts and prototype · research plan, script, recruiting, and 10 moderated sessions · synthesis · final specs, edge states, admin settings, mobile · design QA

Overview

TalentReef is hiring software for restaurants, hotels, and retail chains. The brief was duplicate applications. The data said managers had bigger problems: applications with nothing in them, history they couldn't find, and a score nobody trusted. I took the bigger problem.

The impact

What shipped: an AI summary that cites the source of every line, a new All Applications tab with no AI in it, and a Posi-Fit score that finally explains itself. All 6 customers I tested with saw value in All Applications, and one called it his “magic wand.” Shipped September 2026. In the first weeks, review time roughly halved and managers went from about six tabs per applicant to about two.

Problem

Hourly hiring managers decide whether to move on from an applicant in anywhere from a few seconds to five minutes, usually between deliveries and covering shifts. Opening an applicant in TalentReef gave them a drawer with five tabs and very little in any of them.

Three things kept going wrong:

  • Nothing to read. Fast apply skips work history, so managers saw “N/A” everywhere.

  • History was buried. Checking past applications meant searching a name and opening each one. One manager kept around 100 resume PDFs on his desktop just to keep track.

  • Nobody trusted the score. Strong candidates scored low, weak ones scored high, and nothing said why.

01 · Context split across tabs: Overview, Files, Application, Messages, Notes, History · 02 · Overview opens on contact details and little else · 03 · No sign this person has applied before

The solution

An AI summary that cites the source of every line, an All Applications tab with no AI in it, and a Posi-Fit score that explains itself.

One part for each thing that kept going wrong: something to read, history in one place, and a score that gives its reasons.

Research

Before designing anything, I wanted to hear what managers were actually complaining about. 9,008 NPS responses had plenty to say.

NPS

  • Blank applications were the largest cluster of complaints. Duplicates barely came up.
  • May and June 2026 responses said the same thing, so it was still happening when the project started.
84Complaints about missing application data
2Complaints about duplicates
From 9,008 FY26 NPS responses: 2,566 detractors, 840 flat zeros.
“TalentReef gives NO info to us for the applications we get, all we get is a name and # & address.”
NPS response, FY26 · 1 of 3

Pendo and session replays

  • An in-app poll reached 21,500 managers.
  • 47 session replays matched it: managers jumped between tabs and rebuilt each applicant’s story by hand.
  • AI can’t fill in data that was never collected, so the design had to be honest about empty applications.
What was missing · 40 answered
Seeing prior applications28%
Not switching between sections28%
What they wanted · 37 answered
A quick summary27%
Prior application history27%
One place for everything27%
One poll response set the tone for everything after.
“The solution here is not an AI assistant but rather an improvement on our online application itself.”
Pendo poll · Hiring manager

Competitive analysis

  • I tore down 20+ ATS and AI review tools across duplicate handling, application history, and AI summaries.
  • I scored each against eight principles I wrote for this project.
  1. 01Glanceable and expanded by default
  2. 02In place, never a link out
  3. 03Decision-ready, with the noise cut
  4. 04Prior-application context: why released, no-shows, manager notes
  5. 05At-a-glance flags for multiple and prior applications
  6. 06Actions scoped to the application being viewed
  7. 07Honest when data is missing, never fabricated
  8. 08Within the manager’s scope, but cross-location history visible
Every tool could flag a duplicate. None showed why someone left last time.
20+ products scored against eight principles. Every tool could flag a duplicate. None showed why someone left last time.

What I learned

01

The best tools structured their AI output instead of writing paragraphs.

  • A source behind every claim.

  • “The candidate doesn’t match” kept separate from “we don’t have enough information.”

  • Lever’s strengths and “areas to clarify” shaped my naming later.

02

The biggest gap was prior history.

  • Any ATS can flag a duplicate.

  • None showed why someone was released, whether they no-showed, or what the last manager wrote.

  • TalentReef already had that data.

03

So the brief changed from “detect duplicates” to three jobs:

  • Understand a person from very little.

  • See their history in one place.

  • Understand their score.

Ideation

The early prototype: a strong start, built around duplicates and history. 01 · One AI sentence carrying a verdict · 02 · Counts on every chip · 03 · Everything stacked on the Overview tab, ahead of Files, Application, Messages, Notes and History

The prototype's AI judged the applicant and told the manager what to do. I kept the idea and took away both.

Summary, Posi-Fit, and full history in one card on Overview. Everything in one place, but taller than the screen.

Three ways to write the summary

I wrote the same summary three ways before it worked. Finalized structured method and user testing backed it.

Highlights list. Fast, but strengths and red flags sat in one list, with no sources.

Strengths and Considerations, not weaknesses

  • Call it “weaknesses” and managers start counting pros and cons.

  • A consideration is just something to check, like a schedule conflict, not a reason to say no.

What the AI can actually see

  • Half my first draft described data the model couldn’t read.

  • Every line had to pass three checks: is it captured, is it structured, can the model read it? First casualty: "Checked, nothing to flag." Nothing had been checked.

Testing

10 sessions, all mine: the plan, the recruiting, the script, the moderating. 4 internal users, then 6 customers, from one store to 115 hotels. I recruited through an in-app Pendo guide and picked people by their answers to my open questions, not by who clicked first. Five script drafts later, no task ever asked "would you hire them?" Just "what would you do next?

11 of 14

coordinators forgot by evening what they’d approved

So I designedApprove on the phone, on set

The part where people got excited

01

Cut the time in half

“I think I’d probably cut the time in half, honestly.”

H

Hiring manager

Customer, round 2

02

Magic wand

“This would be my magic wand right here, is the All Applications.”

F

Franchise operator who started out skeptical of AI

Customer, round 2

03

Very little time

“This info is exactly what I look for when I’m vetting candidates, so having this right at the top makes that very helpful, especially for someone like me who has very little time to go over applications.”

H

Hiring manager

Customer, round 2

04

The real data

“The way that the AI is being used now is helpful, because it’s not limiting me from seeing the real data.”

I

Internal user

Round 1

Designs

Who is this person? Have they applied before? Why this score? Each question gets its own part of the drawer, and only the first one uses "AI"

Pick a question to see which part answers it.
AI SummaryUses AI

Strengths and Considerations, each with its source underneath.

Everything below comes from records, not AI
Posi-Fit breakdownNo AI

Up to four factors rated High, Medium, or Low, taken from the company’s own questions.

All ApplicationsNo AI

Every active and past application across locations.

Who is this person? Each line shows its source. Expand it to see what the applicant actually wrote. The ARIES badge marks the only AI on the screen.
Same summary, on the phone most store managers actually use.
When there isn't enough to go on, it says so instead of guessing.
Have they applied before? Every application, every location, active ones first. Expand a row for progress, the latest note, and who to call. No AI.
Full history, still one thumb away.
Errors, empty history, and failed loads each get their own message and a way to try again.
Why this score? Up to four factors, rated High, Medium, or Low, plus what's left out: education, employment history, EEO. The AI never touches it.

Details

Easy to miss in a screenshot. Each one came from something a manager said or did.

It remembers how you left it. Open by default for new managers, then it follows each manager's own habit.

Schedule clashes become a Consideration. An internal tester suggested it, and it shipped as "Not available weekends · conflicts with a posting requirement."

Source

Every line says where it came from. If you can't see the source, you can't check it.

Count

The tab shows how many applications there are, so you know there's history before you click.

Match rule

"Matched by email + phone" tells managers how records were linked, so nobody assumes two people are the same person.

Freshness

"Last generated" tells you how old the summary is. A customer asked for the same thing on notes.

Results

Early data from the first weeks after launch. Figures are approximate to respect client confidentiality.

Every number ties back to a problem from research.

I tied every metric to a problem from research, so a result can’t look good without solving something real.

Blank-application complaints in NPS. Baseline: 84 in FY26, tracked as a share of responses in the next cycle.

Tabs opened per applicant. Managers told us they clicked through 7 or 8.

How often managers expand a source. If nobody ever checks, the summary may be getting too much trust.

How often items get flagged as wrong. This tells us whether the AI stays true to what the applicant submitted.

Lessons

AI can't fill in data that was never collected. A confident summary of a blank application is worse than no summary at all. Same with the score: managers didn't need a different number, just a reason to trust the one they had.

What I’d change

I was too close to moderate neutrally. In one session I caught myself demoing instead of researching. Next time, a second moderator.

What’s next

  • Ask questions about an applicant ("has this person applied to other stores?")

  • Interview questions based on the Posi-Fit breakdown

  • Sorting and filtering in All Applications

"I used AI to cluster the 9,008 NPS responses, then read and synthesized them myself."