Full path.
Case study · Bench · 2026
A comparison that stays a decision, not a six-school brochure.
Six HCI programs, two or three at a time, with the source left on the number.
Challenge: an Indian product designer can collect six brochures and still not know which pair to apply to.
- Role
- Product design and the working build: structure, copy, comparison rules, and the interface.
- Team
- Solo. Shreeraj Narwade. No research participants.
- Timeline
- October 2026. One build, then a reliability pass.
- Platform
- Browser. Program briefs, a compare URL, light and dark themes.
Pittsburgh, Pennsylvania
Carnegie Mellon MHCI
- Visa prior
- 60–75%
- Job security
- 9 / 10
- All-in sketch
- $113,000–$119,000 plus fees
Seattle, Washington
UW MHCI+D
- Visa prior
- 55–70%
- Job security
- 8 / 10
- All-in sketch
- $94,000–$108,000 plus fees
- Why two or three is the comparison, not six.
- What was checked, and what was never asked of anyone else.
- Which directions I did not build.
- What is still unmeasured.
Programs in the set
6
CMU, Georgia Tech, Berkeley, UW, Michigan, UT Austin.
Indices, each 1 to 10
5
Relative to these six only. Visa is a range, not a score.
Programs in one comparison
2 to 3
The URL holds the selection. A fourth pick is refused.
These are counts of the tool. They are not measured outcomes.
/ The problem
The questions were mine, and each one had a cost.
01
Which school is actually for product design?
A brochure line and a curriculum are not the same claim.
Evidence: the six pages do not share a curriculum shape. Judgement: sorting by logo would hide that.
High. It changes which two schools you open first.
02
What does the year cost once rent is in the number?
Tuition alone makes a cheap school look cheaper than it is.
Evidence: bursar tables and labeled living bands, checked 11 October 2026.
High. It can flip a one-year program against a two-year one.
03
Will an F-1 and a design portfolio survive the search?
A studio that does not sponsor is a different plan from a PM seat that might.
Assumption: studios sponsor less often than PM and engineering-adjacent roles. Not a sample.
High for this reader. Irrelevant if you do not need a sponsor.
Writing another ranking was not the problem. Holding two schools to the same fields was.
How might one reader compare six programs without pretending the list is a quality ladder? Constraint: sourced numbers, two or three at a time, no accounts.
01
Six tabs
Each school sounds like the right one.
02
A spreadsheet
Official and estimate share a cell.
03
The stall
No pair is chosen, so nothing is compared.
04
The ask
Same fields, two or three, source on the number.
/ Who
One reader, three pressures that argue.
Assumption, not a study: the reader is an Indian product designer on an F-1 budget. Portfolio, cost, and visa pull apart. A school that wins on craft can lose on rent.
The brand will carry a thin portfolio.
Hypothesis: in-state rumors are the price.
The nonresident line is the price that matters here.
Hypothesis: any design job sponsors.
The visa note treats studios and PM seats as different priors.
No interviews. These lines are hypotheses I used to write the pages.
/ Research
No interviews. Three checks, labelled as checks.
- 01Read the briefn = 1, the authorPRD, information architecture, indices note, and the university table. October 2026.
- 02Check the numbersProgram and bursar pagesEach figure on a program page is marked official or estimate, with a link and a date.
- 03Heuristic passThe built interfaceAfter the first version: card contrast, empty icons, a silent three-program cap, and a one-school compare that looked finished.
/ Insights
What I treated as true, and where it went.
Select a row to mark the chapter that answers it. The table stays readable if you never do.
Finding. Six schools at once stop being a decision.
Evidence. Judgement, from the brief. Not a timed test.
Principle. Compare two or three, and say why the fourth is refused.
Finding. A tuition line without rent, or an estimate without a label, both mislead.
Evidence. The university table and the program pages. October 2026.
Principle. Mark official and estimate, and keep the link.
Finding. A numbered list will be read as a quality ranking unless the page says otherwise.
Evidence. Hypothesis, then the hero card copy.
Principle. Name the weighting in the same view as the order.
/ Alternatives
Three directions. One was built.
Long ranking page
Every field for all six, in one scroll.
Pro: nothing is hidden. Con: the reader never has to choose a pair, so the page becomes a brochure again. Complexity: low. The page is one template.
Spreadsheet export
Rows, filters, no essay.
Pro: easy to sort. Con: official and estimate collapse into one cell, and the weighting disappears. Complexity: medium, and the voice is gone.
Chosen
Briefs plus a short compare
A page per school. A compare that stops at three. The URL is the share.
Pro: the decision fits one screen. Con: a reader who wants all six side by side makes two passes. Complexity: the URL has to stay honest.
/ The decision
I kept the cap, and I kept the label on the number.
Judgement. Feasible in the existing routes. Accessible if the refused card stays readable. Maintainable because the cap is one function, not a new account system.
Tried · earlier
A compare control that opened one school and looked complete.
Chose · shipped
The label matches the unfinished state. The first school stays selected.
Because · judgement
A comparison that cannot compare is a broken promise. The cap of three stays. Adding the next school does not clear the first.
/ Shipped and scrapped
What is in the build, and what I left out.
Shipped
Prototype scope decision
Six briefs and a 2 to 3 compare
Why. The brief asked for both.
Trade-off. Depth on six, not a catalog of twenty.
Kept. A reader can finish a pair.
Shipped
Prototype scope decision
Official and estimate labels
Why. Some numbers are published. Some are bands.
Trade-off. More words on the page.
Kept. A citation does not look like a guess.
Simplified
Prototype scope decision
Indices as judgements
Why. No model with a sample behind the 1 to 10.
Trade-off. They can be argued with, and they should be.
Kept. The methodology page says the limit out loud.
Postponed
Prototype scope decision
More schools
Why. The set was fixed at six.
Trade-off. A reader with a different list gets no row.
Kept. The comparison stays short.
Scrapped
Prototype scope decision
Accounts and live scraping
Why. Scope decision in the brief, not a research finding.
Trade-off. Nothing updates itself.
Kept. No sign-in, and no stale scrape presented as live.
/ What shipped
The pages do the arguing. The compare only lines them up.
Chapter 01
Stop at three, and say so.
Selected cards stay full contrast. The ones you cannot add stay readable, with one line: three is the maximum. Remove one to swap.
Chapter 02
Leave the source on the number.
Cost lines carry official or estimate. The methodology page lists what I refused to invent, including acceptance rates only CMU printed.
Considered · not built
One number, no label, so the table looks clean.
Chose · shipped
Official or estimate sits on the same line as the figure.
Because · evidence
The university table mixed published tuition with living bands. Hiding that mix would make a guess look like a bursar line.
Chapter 03
Say what the order is not.
The hero card is a decision order for product design from India. The decision framework is the same feature, named once, for when that weighting is wrong.
- Home. The order, then a door to each brief.
- Brief. One school, with the source still attached.
- Compare. Two or three chairs. No fourth.
/ Outcomes
I have not measured whether this changes a real application list.
There is no sample, no diary, and no before-and-after. The tool holds the sources I checked on 11 October 2026. That is the evidence. It is not an outcome.
Not measured yet
- Time to a two-school compare: one sitting, five readers, stopwatch, from the homepage to a copied link.
- Whether the order changes the schools someone actually applies to: ask two weeks later, n = 5, same brief.
- Whether visa prior is understood without the word list: ask for a definition, n = 5, after one program page.
/ Access and limits
What I checked, and what is still thin.
- Keyboard: skip link, programs button with expanded state, compare tabs with arrow keys, focus ring in both themes.
- Icons: theme and menu SVGs are 16px, not squeezed by button padding.
- Invalid compare URLs explain the fallback instead of dropping the value in silence.
- Small screens: the header collapses before Decision Framework sits on the theme button.
- Gap: I have not sat with a screen reader user. The names are in the markup. That is not a test.
- Gap: copy to the clipboard depends on the browser allowing it. Failure is stated. Success is not faked.
- Privacy: nothing is sent to an account. The compare selection stays in the URL and in local storage on this browser.
/ Reflection
Three things I would not repeat.
- Card titles that follow the page color. On a cream card in dark mode, the title disappeared.
- A disabled control that only fades. It looked broken. The sentence fixed it.
- Three names for one idea. Decision framework is the name. The order is what that framework produces.
Next: five readers, one sitting, time to a copied two-school link. If they cannot say what the order is not, the hero sentence failed.
How I built this with AI
- 01 Context. The PRD, the Clay tokens, the indices note, and the university table.
- 02 Explore. I used a model to draft structure and copy, then cut claims that were not in a source.
- 03 Build. This site: routes, the compare URL, and the program data file.
- 04 Review. A UI pass on icons, contrast, and the three-program cap. That pass changed the card title color and the compare copy.
- 05 Ship, then back. Review sent me to the card styles, not to a new concept.
Review points back at explore. The loop is the card color, not a second product.
Prompt I worked from: fix the cream card where the title matched the card, and remove every em dash. I kept the card color and changed the title to the card's own ink.