(Case study) Higher education・Polish 3D / Polish AI / Polish+
From an in-person event of about 60 students to an online expo with more than 12,000 student logins
Four years of collaboration, from 2022 to 2026. The overseas education expo moved from an in-person event of about 60 participating students to an online expo with more than 12,000 student logins. Information on 289 exchange schools is extensive and complex, so the site presents it on a 3D world map and carries the details in a one-page site.

(AI assistant)
Ask about selection, compare schools, get a report from AI
Students enter a question and choose schools and topics. The AI assistant turns the designated data into a school comparison report, with a source for every line.
(AI assistant)
Answers use only the designated data, and every line cites a source
Student's phone
- Enter a question
- Choose schools and topics
Development team's service
- MaskingStudent IDs, emails and phone numbers are masked before sending
- Designated dataNo content from the web
- Line-by-line checkCitations that do not match are removed
- QuotaWeekly quota and daily cap are adjustable
Model provider
- Language model
Never sent outStudent IDs, emails and phone numbers.
Scope: the question and report flow of the on-site AI assistant.
Related method
How citations are kept honest: any point that fails verification is dropped from the answer.
- Retrieve and compare to the thresholdThe model is involved only once the confidence threshold, which takes in score and coverage, is passed
- Model answers from the sourcesEach point carries a source number
- Validator checks rule by ruleNumbers must appear in the source or the question; the model may not add words of judgment
- Retrieval fails the thresholdReply “no organized answer yet”List the three closest questions, log one miss, hand off to a person
- Validation passes in fullAnswer with sourcesEvery point traces back to a source
- Some points fail validationRetry once, drop the points that still failIf all are dropped, no forced answer; if both attempts fail, return an error and do not count the attempt
(Challenge)
Extensive information, and no easy way to show how many choices exist
The 289 exchange schools are spread across every continent, and the information is extensive and complex. An ordinary one-page site struggles to show that many schools at once, and the page rarely looks tidy.
The expo aims to convey the information simply and clearly and to draw students to apply for exchange. The core message challenges an assumption in one line: everyone has a chance to go on exchange.
(Approach)
A 3D world map shows every school, and a one-page site carries the details
Many online exhibitions now use VR, which suits a sense of space. Finishing a full VR exhibition takes many steps, though, visitors get lost easily, and the content is mostly images and video, which suits large amounts of text poorly.
The home page is therefore a 3D world map that shows how many schools there are at a glance. Complex information goes into a one-page site that is easier to read. A map also invites students to explore independently and makes the visit feel more like walking an expo floor.
On the front end, Three.js renders the 3D models in the browser, and parameters such as swipe speed, map size, boundaries and damping were tuned by hand so that controls behave as expected on every device. To protect performance, the scene uses no real-time lighting. Lighting is baked into textures, shadows under the map tiles are image files kept as small as possible, and the look carries a slight pixel-art feel.
Desktop
The whole world first, then step by step into each continent
The home page opens with clouds and paper planes. Pressing Start exploring flies the camera into the 3D world map, and a six-step tutorial covers the world map, finding schools, application requirements, senior students’ experiences, the AI assistant and favorites. After that, the scroll wheel zooms and dragging pans, flying from the full view into each continent, where schools stand as small houses on the map tiles. The menu gathers programs and live streams, events, announcements, the calendar, the school catalog and partner institutions.

Phone
Just as complete on a phone, with the tutorial in a bottom panel
On a phone the map uses one-finger dragging and two-finger zoom, the tutorial appears in a bottom panel, school search stays fixed at the bottom of the screen, and the menu opens from the top right.
Related method
The loading order a map site follows on a slow phone connection.
Before interactive
- Main scene and key partsThe reveal condition waits only for these
- Island shellsPositions come from constants; a placeholder holds the place until the model arrives
- Images on the first screenLayout reserves space by width and height
After interactive
- Models of three islandsAbout 2.5 MB in all, swapped in and faded in on arrival
- Images seen only laterNo URL is given before interactive
- Pages that need no 3DThe countdown page loads no 3D JavaScript
Never block the first interactionWhatever cannot be seen is deferred and added one piece at a time once the user can act
An event-map website, adjustments measured on an emulated phone with slow 4G
(Key design)
Three key designs
Entry page and ordering
The world map is the star, and an entry page carries the full event information before visitors reach it, giving the step into the map a sense of ceremony. The order puts the map on top, then the core message, then the application schedule to raise conversion, and the points game last.
School information
Points and raffle pages are tucked into a modal instead of sitting on the map page. This separates the student-only activity from the rest of the content, and visitors who are not students feel little gap. School information can be explored on the map or browsed as a list with filters, and school quizzes in the modal help students learn about each school.

Points and invitations
Prizes draw students to log in and play the points game. Daily logins, exploring expo content, watching the live broadcast and completing a survey all earn points, and a leaderboard adds competition. The top scorers win prizes such as an iPad or an Apple Watch. An invitation system rewards each successful referral, so more invitations mean more rewards.
(Back office)
During the event, the back office starts with what needs handling
After login, the first thing shown is what needs handling, followed by trends and upcoming sessions. Each metric opens to its full calculation, using the same definition as the data queries.


For fairs and admissions with heavy information that need AI answers, the first meeting can confirm the data and scope.
(Result)
Over four years, participating students grew from about 60 to more than 12,000
From 2022 to 2026, the expo moved from an in-person event of about 60 participating students to an online expo with more than 12,000 student logins. In one edition, 2,000 students logged in with student email accounts within 7 days of opening. Site views for all of the previous year totaled 100,000, and that edition passed the figure in just 10 days. Live broadcast views also set a new record.
More than 20% of logged-in students browsed the site for over 7 consecutive days, the longest single visit lasted 3 hours, and more than 10% of students stayed over 2 hours in a single visit.
School roofs and bodies are designed to suit the tone of each continent, and varied combinations keep the scenery from feeling repetitive. The opening pairs a soaring flyover sound with an original score, carrying visitors from the whole world to each region.


(Field notes)
Methods and measurements, open each one
Front-End Build of an Online Expo: A 3D World Map and Information Architecture for 289 Schools
The display format follows the content: a map plus a one-page site when the information is mostly text and listsThe server provides the school lists and the leaderboard, and the camera, gestures, and search all run in the browserThe map shows the whole picture and helps find schools, while details and events sit on the one-page site and in windowsAfter an AI Answering System Goes Live: Refusal Thresholds, Abuse Guards, and Keeping Answers Trustworthy
Every path that calls the model passes through the same guard and allowance first Related method
User’s browser
- QuestionThe user types a question
Development team’s service
- Entry guardProgrammatic matching for injection, abuse, and gibberish, plus allowance and rate limits
- Retrieval and thresholdNo model call if the threshold is not met
- ValidatorChecks the model’s draft rule by rule
Model provider
- Language modelDrafts an answer from the sources
Not sent outA question that hits a guard never reaches the model; the log keeps only the category, not the original text
Scope: the user-request path of the answering service; batch work uses a separate key
Evaluating an AI Q&A System Before Launch: Fixed Test Sets, Confidence Thresholds, and Abstention
In evaluation, model calls become files, and the same request always gets the same answer Related method
Q&A system
- Break down the questionEach call gets a fixed file name derived from its request
- Write the answer and verifyAnswers that fail verification are moved to a separate folder with a failing exit code, and reappear in the next round
File directory
- Request fileOne per call that has no answer
- Answer fileAn existing answer is reused, never re-answered
Stand-in answerer
- Small modelOne request per run, answered question by question
Stays inside the boundaryThe database is read-only; API-key environment variables are cleared in evaluation mode; network requests to the model provider are blocked.
Scope: offline evaluation mode, with no real model and no external services.
Once a question arrives, retrieval score and verification results route it to an answer, an abstention, or a person Related method
- Retrieve and scoreRetrieval score and fit value computed from user phrasing, keywords, and answers
- Compare to thresholdsBoth values must clear the thresholds
- Verify each claim's sourceSource number, numbers, and tone must all pass
- Score clears the threshold and every claim passes verificationAnswer with sourcesEach claim carries a source number
- Score is below the thresholdDecline and hand off to a personState that no prepared official answer exists yet, list the three closest questions, and log a miss
- Some claims fail verificationRetry once, then drop claims that still failKeep only the claims that pass
- Every claim has been droppedDo not force an answer; hand off to a personNo answer is produced
The client writes questions and labels answers, the development team scores and tunes thresholds, and both review before launch Related method
The client establishes the test set; the development team measures thresholds against it.
RAG for Enterprise Knowledge Q&A: A Verbatim Verification Workflow
About a tenth of the excerpts the model picks fail the verbatim check Related method
- Long PDF documentsabout 1,600
- Model picks excerptsabout 15,000
- Code compares to the sourceNormalize, then locate by string search
- Dropped on mismatchabout a tenthIncludes those under 10 characters
- Passedabout 13,000
Rounded figures from one full run.
A low score or a fully stripped answer goes to a person instead of a forced reply Related method
- Keyword scoringBM25 plus synonym expansion
- Compare to thresholdMeasured from test questions
- Check each pointNumbers must be in source or question
- Score falls shortList closest questionsHand off to a person
- Check failsRetry once with the reasonDrop failing points if it still fails
- All points droppedReply “not ready yet”Hand off to a person
The flow in the customer-service version.
Measuring Web 3D Performance: Common Errors and Methods That Hold Up
On the same scene, a single render looks several times more expensive with vertical sync off than with it on Related method
- about 20 msTime per render after entryUse the sync-on figureabout 1 to 6 ms
First 40 renders after entry in a 3D map, approximate values; vertical sync on means a true 60 Hz
(Client note)
In this collaboration, POLISH showed exceptional design capability and mature professional judgment. The website's overall structure is clean and sharp, with well-judged white space, clear and orderly information, and smooth operation that gives users a fluid, comfortable browsing experience. The visual style is calm and minimal yet rich in texture, striking a fine balance between function and aesthetics. The design goes beyond surface appearance, with care visible in the details and a strong focus on user experience. Overall, it never strains to impress, yet attention to detail shows everywhere. It is a mature work in which technical expertise and aesthetics coexist.
DavidPM・NTUOIAThe first meeting takes 15 to 30 minutes and gives a direction, a budget range and a draft timeline.
Polish 3DPolish AIPolish+Next case: A 3D scene explains the wind-power service range, with each service shown online during sales meetings
(Contact)
Start a project.
Three steps to a clear direction and budget range.
- 01Book the first meeting: choose a time. A full specification is not required.
- 02A 15 to 30 minute online consultation: direction, budget range, and draft timeline.
- 03After materials are reviewed, a quote and delivery specification follows in about 3 to 5 working days.
- hello@polish-design.com.tw
- Mailing address
- 3F, No. 191-1, Zhifu Rd., Zhongshan Dist., Taipei 104, Taiwan
Book the first meeting
Monday to Friday, 16:00–18:0030 minutes online
Choose a time, then confirm with a Google account or an email address. A confirmed booking can be added to the calendar.
Choose a time



