Carrington SmurlProduct Designer / Engineer

AI Product Design2026

Saddle Pass

I ran the research, wrote the specs, designed the screens, managed the design team, and built the agents that checked our output. This is the case study where all of those turn out to be the same job.

Role
Founder · Lead UX Designer · Design Manager
Timeline
May 2026 – present
Team
3 engineers, 1 design director, 1 contract designer, 2 design interns
Surfaces
Rider · Barn owner · Horse owner · Service provider · Admin
Stack
Figma · React / React Native / Flutter · Design tokens with tests · Claude Code
Constraints
  • Four user types whose real-world identities overlap
  • Two sides of a marketplace with the same complaint and opposite incentives
  • One design system consumed by three codebases
  • Pre-launch — discovery only, no usage data

The problem

The horse world still runs on Facebook groups, text threads and paper calendars.

That is one sentence, and it took four months of research to earn it. Four groups of people are all stuck in the same gap, for different reasons:

Riders Finding a lesson or trail ride means scrolling Facebook groups and waiting days for a text back.
Barns & trainers Bookings, boarding waitlists and payments live in notebooks, spreadsheets and Venmo.
Service pros Farriers, vets and bodyworkers grow by word of mouth only. There is no place to get found.
Hiring Barn jobs fill through who-you-know. A bad hire costs horses, clients and money.

And underneath all four:

The missing piece is trust: no vetting, no shared reviews, no protection when something goes wrong.

That last line is the one that decides the product. A marketplace that only fixes discovery in this market is a better Facebook group. Trust is the thing nobody has solved, and it is a design problem before it is a feature list — it lives in verification states, in what a stranger can see about you before they book, and in who is allowed to do what on a barn’s behalf.

It is also a market with no digital home rather than a market with bad software. The core customer is the adult amateur rider plus the barns, trainers and pros who serve them, across every discipline from Western to dressage, and Texas alone is a multi-billion-dollar horse economy with a few hundred thousand horse owners in it. That scale is why the answer is a marketplace rather than a tool.

Research first, because the problem was not obvious

Everything above is stated with a confidence I did not have at the start. Plenty of markets look broken from outside and work fine for the people inside them, and the horse world is proud, close-knit and perfectly capable of preferring its group texts. So the question was never whether I could see the gap — it was whether the people in it wanted it closed.

So before any screen, I fielded a 47-question discovery instrument to horse professionals, and a second cut aimed at the business side.

The instrument is the part I would defend in an interview. It maps the current state before it goes anywhere near a solution — how clients book today, how cancellations are handled, which admin tasks eat the week, how payment gets chased. Only then does it ask about the product, and the two most useful questions are the uncomfortable ones:

  • Q34 — “What would make you NOT want to use a marketplace/booking platform?”
  • Q35 — “What fee structure would feel fair?”

Most founder-run surveys ask whether people like the idea. Asking for the objection and the price up front is how you get something you can design against instead of a compliment.

Both sides described the same failure

Providers, on booking: “It’s not centralized.” · “Not having everything in 1 place.” · “Keeping tabs on who, what, when.” · “Taking away the need for the back and forth communication of booking.”

Riders, on booking anything horse-related: “Lots of back-and-forth messages.” · “Hard to find pricing.” · “Don’t hear back from barns.” · “Not sure what’s safe or reputable.”

Supply and demand naming the identical failure, unprompted, is the strongest signal a two-sided marketplace can get in discovery. It means one product can serve both without trading them against each other.

The objection that shaped everything after

Q34 came back overwhelmingly consistent. Nearly every respondent said “fees too high.” A clear majority said “too complicated to learn.” Three said “takes control away from me.”

That ordering directed more of the design than any feature request did. A platform these users adopt has to be visibly simple, and it must not feel like it takes the barn out of the barn owner’s hands. That last objection is not a pricing problem — it is an interface problem, and it returns later in this case study as a permissions system.

The product, in three words

The whole thing is organised around a sequence rather than a feature list — Look. Book. Ride. — because that is the order the user actually moves in, and it keeps the team from arguing about navigation.

Look Search lessons, trail rides, boarding, services and jobs by discipline, distance and price. Every provider is vetted.
Book Request, sign the waiver and pay in one flow. No phone tag.
Ride Day-of check-in, reviews and one-tap rebooking keep riders coming back.
The request-to-book screen for a private beginner trail ride with Hill Country Riding Co.: a booking details card, a starting price of $85, a May calendar with selectable dates, and a preferred-time field.
Book — one flowRequest, waiver and payment in a single pass. “No phone tag” is the design requirement hiding in that sentence: every step that bounces a rider back to a text message is a place the booking dies, and the old way was nothing but those steps.

Five pillars, one account

Profiles & accounts · marketplace · job board · boarding · events. One account covers all of it, because a person in this world is rarely only one thing.

The rider onboarding step: riding preferences with chips for booking trail rides, taking lessons, finding boarding, horse care, buying or leasing, listing services and hiring help; a ride-frequency selector; and a checkbox for signing up a child under eighteen.
Onboarding — riderThe preferences a rider sets are the same taxonomy the marketplace searches on, so onboarding is doing double duty as the first act of personalisation. The child-under-18 checkbox is the one that matters operationally — minors change the waiver, the insurance and who may book.
The provider onboarding step: riding styles and disciplines, with a dense grid of selectable chips including racing, dressage, eventing, western pleasure, reining, barrel racing, natural horsemanship, ranch work, polo and working equitation.
Onboarding — service providerThirty-odd disciplines as chips rather than a dropdown, because discipline is how this market actually sorts itself and a provider who cannot find theirs assumes the platform is not for them. Getting this vocabulary right was research output, not copywriting.
The horse-owner onboarding step: your horses, with an add-horse-manually button and a bulk-upload area for registration, breed registry, pedigree, Coggins and vet records, noting that in this beta build the files are held and wired up later.
Onboarding — horse ownerBulk paperwork upload with an honest note that the beta holds the files rather than parsing them. Horse records are the owner's real burden — Coggins, registration, vet history — and a product that asks them to retype it will lose to the folder they already have.
A horse profile for Atlas, a nine-year-old Quarter Horse, marked draft with availability pending, showing empty care-and-health-notes and documents sections and a booking-foundation panel listing availability not set and a platform waiver placeholder.
The horse as a recordA horse is a first-class object here, not a field on a person. The “booking foundation” panel is the honest part of the design: it states out loud what is still missing before this horse can be booked — availability not set, waiver placeholder — rather than letting an owner think they are live when they are not.

We monetise supply, not riders

The decision that shaped more of the interface than any visual choice: the barns and providers pay, and the riders do not.

That inverts the usual marketplace instinct, and it means the product has to earn its keep on the supply side by replacing the back office these businesses have never had — the day sheet, availability, boarding waitlists, payouts. Those are not secondary admin screens. They are the product the paying customer actually opens every morning, and they are why the permissions model further down this page is load-bearing rather than a settings page.

Research into specification

Findings only count if they survive into the build, so the specs live in Figma beside the screens rather than in a document nobody opens.

A product requirements document laid out as a Figma frame, titled PRD · Events v1.0, with structured requirement sections.
PRD · Events — v1.0One of six PRDs I wrote as Figma frames, versioned and dated, on the same page as the screens they govern. Putting the requirement a few hundred pixels from the artwork is the cheapest way I know to stop a spec and a design drifting apart — nobody has to go looking for it.

Managing the work, not only doing it

With a design director, a contract designer, and two interns in the file, the structure of the file is the management system. Pages sit in four bands:

  • DELIVERY — ✅ Dev-Ready (Final Drafts) · 👀 Review-Ready · ↔️ Responsive
  • WORKING — 🔵 Working (carrie) Source of Truth · 🦄 Future Features Phase 2
  • RESEARCH — Discovery · User Research
  • ARCHIVE — 💀 Archive/Outdated

Every delivery section carries the date it reached that stage — DEV-READY (13 Sep), REVIEWING (31 Aug). An engineer opening the file never has to ask which version is real, and a designer never has to ask where to put something.

Two smaller conventions do a lot of work. 🚩 REBUILD FLAG marks a component that shipped wrong and must be redone, so the debt is visible in the file rather than in somebody’s head. ⛔ copy-gated marks the insurance and protection screens, where the words cannot ship until a partner and a lawyer sign off — so nobody builds against copy that is going to change.

A/B testing my team’s work against my own

The thing I am proudest of here is a critique method, not a screen.

Saddle Pass ran two parallel design tracks: a dashboard track owned by another designer, and the consumer track I owned. Rather than argue about which was better, I ran a structured comparison.

A dark green key frame explaining the A/B method: Sanjay's Dashboard V3 screens versus our Consumer V3 screens, dimension-by-dimension verdicts, colour coded blue for his wins, green for ours, gold for ties or merges.
A/B test keyThe method stated up front and applied identically to both sides. Dimension-by-dimension verdicts across eight areas — information architecture, the edit-profile pane, role composition, team and permissions, public pages, modals, system density. Blue where his won, green where mine won, gold for a tie. Then a refurbished screen taking the winner's structure with the loser's best ideas grafted on.

Three things make that worth more than a design review:

  1. The dimensions are fixed before the comparison, so a verdict rests on a stated basis rather than on whoever talks last.
  2. Losing is visible. Blue appears on that board. I did not win every dimension and the record says so.
  3. The output is a merge, not a verdict. Each chosen screen is marked ✅ CHOSEN and is a synthesis.

The third is the actual management skill. Picking a winner takes thirty seconds. Taking the losing design seriously enough to find what was right in it is what keeps a designer on your team after you have overruled them.

The chosen Edit Profile screen from the A/B round, combining the structure of one design with details carried from the other.
✅ CHOSEN · R1 · Edit ProfileThe merged result — my structure, with the dashboard track's baseline carried through where it was stronger. Marked CHOSEN in the file so there is no ambiguity later about which version engineering builds.

The permissions model, where the research lands

Remember the objection: “takes control away from me.”

A barn runs on staff — barn hands, trainers, a vet, a working student. If the platform cannot represent who may do what, the owner either hands over their own login or does not adopt it.

The Team and Permissions screen with an Edit permissions modal open for a barn hand: a role selector set to Staff — day-to-day operations, and granular capability toggles for bookings and calendar, horses and herd, boarding and contracts with nested sub-permissions, messages, and money set to view only.
✅ CHOSEN · Team & PermissionsA role sets a sensible default — 'Staff — day-to-day operations' — and every capability under it can be customised, down to nested sub-permissions on boarding. Money defaults to view-only. This answers the research objection directly: delegation without surrendering control, and the owner can see exactly what they granted.

This is the same problem I spend most of my time on in AI products. A permission model is a design surface: what may this actor do alone, what needs approval, and how is that made legible to the person who is accountable. This is the barn-staff version of the question I answer for every agent I build.

Four roles whose identities overlap

Saddle Pass onboarding step 1 of 8: three role cards — Rider/Participant, Horse Owner/Trail Guide, and Service Provider — each with a Choose button, under a progress bar.
Onboarding, step 1 of 8In the real world these identities overlap constantly — a horse owner who leases out trail horses is also a service provider. The subhead removes the stakes from the choice: 'Choose the closest fit for now. If you own horses and provide services, we will capture both on your profile.' The taxonomy the database needs is not a taxonomy the user has to commit to.
The member profile screen, version 3, marked dev-ready.
6.1 Member Profile — V3 · DEV READYProfile carries the marketplace's trust problem. Riders told us they were 'not sure what's safe or reputable'; this is where that gets answered, through verification state, history, and what a stranger can see before booking.

The assistant, scoped by copy

Ask Saddle Pass AI screen: an AI avatar, a Beta assistant pill, the heading 'Ask an equine question', a question field, an Ask Saddle Pass AI button, and three quick-start prompt chips.
Ask Saddle Pass AIThe placeholder scopes the domain out loud — booking, safety, horse records, barns, services — so the boundary is stated before the user invests in a question. The 'Beta assistant' pill sets the accuracy expectation where it will actually be read. The quick starts solve the blank-input problem and demonstrate what a good question looks like here.

Scoping the assistant was a product decision disguised as a copy decision. That placeholder is doing the work a refusal message would otherwise have to do later, at a worse moment, after someone has typed out a question about their sick horse.

Where AI actually entered the process

Not as a feature. As how the work got done.

I design in code as well as in Figma. Working in Claude Code from the terminal collapsed the distance between a decision and a running implementation from a sprint to an afternoon. The specs I handed engineering were ones I had already built against.

The design system is enforced by tests, not vigilance. Tokens live in the repository with a sync script and tests that fail the build when any of the three codebases — React, React Native, Flutter — diverges from source. A system nobody checks drifts within a sprint. This one cannot.

I built agents to QA the work. Screen capture, review sweeps, and comment triage run as automation rather than as somebody’s afternoon. The design-capture parameters in the web app — render a single view, suppress the entrance animation — exist so an agent can produce a clean, repeatable screenshot set with no human driving a browser.

That is the throughline between the design degrees and the engineering: research says what to build, design decides what it should be, and automation is how a small team holds the quality line once it exists.

Where it stands

An earlier build of the home screen with a green branded header bar carrying the Saddle Pass logo and a Beta chip above the greeting, search field and category chips.
Before — branded chromeThe earlier home screen wore its own logo. On a phone, in a barn aisle, that bar costs a row of content to tell a signed-in user the name of the app they just opened.
The current home screen, which drops the branded header and opens directly on the greeting, search field, category chips and featured rides.
After — straight to the workThe same screen with the chrome removed: it now opens on “Hi Carrington”, the search field and the rides. Branding earns its place on a marketing page and spends someone's attention everywhere else.

The design is built and handed to engineering across rider, barn owner, horse owner, service provider and admin surfaces, specified against the token system. The web app is complete and gated behind a feature flag, waiting on beta.

The near-term plan is sequenced rather than scattered: the EquiStaff site, job board and AI hiring assistant go live first, then the Saddle Pass MVP beta in Austin with the first ten barns and providers each taking real bookings, then in-app payments switched on, then provider subscriptions, then the iOS app through TestFlight. Partner work funds the platform while the platform gets built, which is the whole reason the studio and the marketplace sit under one roof.

What I will not claim

There are no outcome metrics, because it has not launched to users. Anything I told you about conversion or retention would be fiction. What I can defend is the chain: a 47-question instrument, a quantified objection set, specs written against those findings, screens A/B-tested against a second designer, and a token system with tests that stops it drifting on the way into three codebases.

What I would tell you in an interview

I did four jobs here and they were not separable. The research produced the objection list, the objection list produced the permissions model, the permissions model is the thing a barn owner adopts or doesn’t, and the token tests are why any of it survived five contributors and three codebases.

What I would do differently is measurement. I have a defensible rationale for every screen and no instrumentation plan written down before handoff. Deciding what you will measure is easiest while you still remember why you chose the thing — I left it until later, which is how it ends up not happening at all.