Tailoring a resume should take a minute, not an evening.
Most applications go through software before a person. Refynes rewrites the resume for each posting and shows the score, so people apply with a resume made for that job.
AI product · web, iPhone, Apple Watch
Refynes
Paste a job. Ryn tailors the resume.
An AI resume app I designed and built: a companion called Ryn rewrites a resume for each job, shows the ATS score, and runs on one glass design system in light and dark.
- Role
- Designer and builder
- Scope
- Product, design system, front end, AI flows
- Focus
- AI transparency, ATS clarity, light and dark
- Role
- Designer and builder
- Product strategy, UX, the design system, the React front end, the iPhone and Apple Watch apps, and the AI prompts behind Ryn.
- Focus
- AI resume tailoring
- Paste a job posting, get a resume rewritten for it, and see how an applicant tracking system will read it.
- Platform
- Web, iPhone, Watch
- refynesapp.com, an iPhone app on the App Store, and an Apple Watch companion for interview reminders.
- Year
- 2025–2026
- Built and shipped continuously; the glass design system and Ryn's tailoring card landed in late 2026.
- Duration
- Ongoing
- A live product: changes ship weekly, from copy to the AI model behind tailoring.
Project arcOngoing - Core loop, completed
- Ryn, completed
- Glass system, completed
- Measure, completed
I designed it, built it and shipped it.
From the first prompt to the paywall: the flows, the Ryn character, the light and dark glass system, the React app, the iPhone and Watch apps, and the AI calls through the Anthropic API.
Ryn at work
The wait became the explanation.
Tailoring takes several seconds of AI work. It used to be a looping animation. Now Ryn’s card shows what is happening: reading the job, matching the experience, rewriting the lines, the keywords it finds and the score as it climbs. Then every edit plays back, one at a time.
Keywords appear as Ryn finds them, coral when missing and mint when matched. The score ring moves with the work, not after it.
Cancel used to sit in a separate bar at the bottom of the screen. It moved inside Ryn’s card, beside the progress, and cancelling leaves the resume as it was.
Each rewrite plays back with the old line struck through and the new words highlighted, so users see exactly what the AI did to their resume.
Score before you apply
Show people what the robot reads.
An ATS score is only useful if people can see why. One tap switches the resume to ATS View: formatting stripped, in monospace, the way a parser sees it. Styles say whether they are safe before anyone picks one.
Monospace, no color, no columns. Headers a parser may not recognize are flagged in yellow.
Two-column styles say “ATS Risk” on the option, before someone applies with a resume a parser may scramble. Ryn’s quick actions put the same checks one tap away.
Design system
One glass, two lights.
The light theme had grown a grey haze, beige panels and solid teal selection slabs. I rebuilt both themes on one token set: the same names point at light values or dark ones, so every component works in both. Then I documented it atom by atom and measured it.
Selected cards are a soft tint with an emerald edge, never a solid block.
No component knows which theme it is in. They only use names like surface.glass and action.primary.
Measured, then fixedFive pairs failed WCAG AA. They now pass, in the app.
Building the library meant measuring every text and background pair in both themes. The fixes went back into the Refynes stylesheet, not just into the documentation.
- Primary button label, light3.20 → 4.54
- Accent text on mint chips3.89 → 4.62
- Error and ATS-risk text4.02 → 5.25
- Placeholder, light / dark3.55 → 4.90 · 4.32 → 6.12
- Focus indicator1.36 → 6.05
Tokens for both themes, atoms through templates, a working Ryn tailoring card, and the measured fixes, as an interactive library.
Off the phone
Interview day is on the wrist.
The Apple Watch companion pairs with a code and does a few things well: a reminder the day before an interview, a two-minute briefing, a pitch to rehearse, and Ryn by voice.
Process
How Refynes came together.
Four overlapping phases. Each one changed what the next one had to solve.
Start from the job posting
The core loop came first: paste a posting, get a resume rewritten for it, see the score. Every other screen exists to make that loop faster or more trustworthy, including the PDF import that reads two-column resumes in the right order.
Give the AI a face
Ryn started as a 3D model, rendered to sprites in Blender. Ryn guides empty states, answers questions and, most of all, narrates the tailoring run, which replaced an abstract loading animation.
One glass, two lights
An approved board, then a sweep of about 490 hard-coded colors onto one accent family and a shared set of glass tokens, so light and dark are the same components with different values.
Document, measure, ship back
The atomic library: tokens to templates, every pair measured in both themes. The five failures were fixed in the library first and then shipped in the app’s stylesheet.
Outcome
The AI shows its work.
What shipped, what is not measured yet, and what I would take further.
What shipped
- Paste-a-job tailoring with a live ATS score, on the web and in the iPhone app
- Ryn's tailoring card with steps, keywords, a moving score and Cancel inside the card, followed by the edit reel
- ATS View and ATS-safety labels on all eight resume designs
- One glass design system for light and dark, documented as an interactive library with five accessibility fixes shipped back
- An Apple Watch companion for interview reminders, briefings and Ryn by voice
What was measured
Refynes is early and its user base is small, so no conversion, retention or hiring numbers are quoted here.
The call I’d revisit
I'd make the edit reel teach, not just show.
The reel shows each edit and lets people keep or skip it. It does not yet say why a change helps: which keyword from the posting it answers, or which ATS rule it fixes. One line of reasoning per edit would turn a review step into something people learn from for the next application.
What I’d instrument next
- Share of tailoring runs cancelled before they finish
- Edits kept versus skipped in the edit reel
- ATS score change per tailoring run