1
Customer3D citySix hotspots lead straight to each page
(Case study) Signage and advertising・Polish 3D / Polish AI
Signage previewed on the real storefront before payment, so clients see the finished sign first
A custom sign is usually approved from a flat drawing. The website became a 3D city where clients upload a logo and a storefront photo, preview the sign lit at night, and carry presentation mode straight into a proposal meeting.
The same approach suits any trade where clients see the product only after it is built: signage, building materials, windows and curtain walls, interior fit-out, exhibition booths and lighting.
5sign styles
9storefront scenes
4demo buildings
(Challenge)
A sign shows its effect only when placed on a building and a street
A sign shows its value only when placed back on a building and a street. Flat drawings and renderings cannot answer three questions: whether it stands out on this building, how it looks lit at night, and whether front-lit or painted lettering suits it better.
Sales staff used to add explanations in words, and clients had to imagine the finished sign. The website was meant to offer a place where the sign can be seen first.
(Approach)
The website itself is a city
Signage company websites usually show finished photos, which convey past work. A client wants to confirm what the sign will look like on the client's own building, so this website answers that with an operable scene.
The home page is a 3D city that can be dragged and zoomed. Six hotspots stand between the blocks and lead to consultation, selected cases, service plans and the sign showroom.
City textures were reduced from 4K to 2K and recompressed as KTX2, cutting GPU texture memory from about 681MB to about 30MB. When a browser cannot run 3D, the page falls back to a static cover and the main links, so no visitor is turned away.
What lets the city run on a phone: shrink the textures first, then switch to KTX2.
- 8 textures at 4096 x 4096about 681MB
- Reduced to 2048 x 2048about 170MB
- Then switched to KTX2about 30MB
GPU texture memory of the home page 3D city. Figures are estimates, and these textures are already in the production release.
(Key design)
Three designs that let clients try it directly
Showroom
Clients choose from four demo buildings and upload an SVG logo, and the sign appears on the facade at once. Five styles switch in one tap: front-lit, back-lit, front-and-back-lit, painted 3D lettering and lightbox. Color, edging, brightness, position, size and thickness can all be fine-tuned. Pressing Present hides the controls and starts an automatic camera orbit, so a laptop is enough for a proposal.
Desktop
Swap logos and styles, and the sign appears on the facade at once
Four demo buildings are available. Logo templates swap in one click, styles switch between front-lit, back-lit and painted dimensional letters, and sliders fine-tune position, thickness and size, with the view updating immediately. In photo simulation, picking a built-in storefront photo places the sign on the facade in the photo.
Photo simulation
Clients upload a photo of the actual storefront and drag the sign into place, which answers the question that matters most: how it looks on this building. Without a photo, nine built-in day and night storefront scenes are available. Photos are processed in the browser only and are never sent to a server.
Phone
On a phone, the sign goes straight onto a storefront photo
Photo simulation works directly on a phone: logo templates and styles switch in one tap, and sliders adjust horizontal and vertical position and up-down and left-right rotation, so the sign lines up with the facade in the photo.
Ease of use
A seven-step guide appears on the first visit to the showroom. The color picker, sliders and upload window all work with a keyboard and screen readers, with no mouse required.
For industries where the customer sees the product only once it is finished, the first meeting can settle how to preview it.
(Result)
Case pages embed an inquiry form, with content in four languages that search engines can read
The site now runs in Traditional Chinese, English, Japanese and Malay. Each case page ends with an inquiry form that asks for the project type and budget range first. Cases also output FAQ structured data and llms.txt, so search engines and AI assistants can read the content.
Explore
Try
Proposal
Inquire
Customer
Sales
2
CustomerShowroomFive sign styles switch in one click
3
CustomerPhoto simulationDrag the sign into place
4
SalesDemo modeControls hidden, camera orbits automatically
5
CustomerInquiry form on the case pageProject type and budget range
Photos are processed only in the browser and are never sent to a server.
New case articles are drafted by AI, with rules taken from statistics on the existing articles.
Set rules
Generate and label
Review
Before launch
Program
Model
Person
1
ProgramCount existing contentImage counts, headings, caption length, FAQ types
2
ModelGenerate and label by the rulesLeave a field empty when no source exists
3
ProgramValidator gateConsistent structure, verbatim match, exit code 0 to accept
4
PersonSupply missing materialA person supplies sources for empty fields
5
ProgramAssemble the review packetReviewers do not dig through the database or source files
6
ModelSample by riskAbout 10%; only hard cases escalate
7
PersonCheck text against images and top-of-page screenshotsPhone and desktop
Summarized from case articles and large-scale document labeling.
(Field notes)
Methods and measurements, open each one
Building a Signage Preview System: Logo Upload, Photo Alignment, and a 3D Showroom
Logos and storefront photos are processed only in the browser, and the server supplies static filesWhen WebGL is unavailable or scene rendering fails, the home page shows a static cover and the content below loads as usualA customer moves from the 3D city to the inquiry form, and sales can use the showroom screen for proposalsQuality Gates for AI Batch Content: Measure Existing Pages, Then Write Checkable Rules
In two pilot rounds, review cost about 20 times as much as labeling Related method, figures from another project
Labeling costReview cost
- 1xMost expensive model reviews every documentPilotabout 20x
Unnamed project; costs are list-price conversions, not invoices; labeling cost is set at 1x.
Only two kinds of hard cases escalate to the strongest model; the rest stay in the sample review Related method
- A program builds the review packetReviewers do not dig through the database or source files
- A mid-tier model samples by riskAbout 10%; long documents, scanned files and doubtful layouts first
- Boundary or coverage score is 3 or below, and the cause is on the labeling sideEscalate to the strongest modelHard case
- The validator failsEscalate to the strongest modelExit code is not 0
- Anything elseStay in the sample reviewValidator passes and scores meet the target
About 1,500 documents were labeled, and model review sampled only about 10% Related method, figures from another project
- Full-run labelingabout 1,500 documentsOne sub-task per document; accepted only at exit code 0
- Not sampledabout 1,350 documentsNo model review; the validator is the gate
- Sampled by riskabout 160 documentsAbout 10%; long documents, scanned files, doubtful layouts and new sources first
Full-run result of an unnamed project, in approximate numbers; documents not sampled still had to pass the validator.
Memory Budgets for 3D on Phones: Which Features Stay on Mobile and Which Are Desktop Only
Texture memory fell from about 681 MB to about 30 MB, at the cost of halving each texture's edgeMeasuring Web 3D Performance: Common Errors and Methods That Hold Up
A number is used only when it comes from a real GPU, after loading, and returns to baseline Related method
- Confirm rendering ran on a real GPUFrame rate and memory from software rendering are rejected
- Confirm loading finished, the tab was in the foreground, and content was drawnNumbers from loading and hidden tabs are rejected
- Finish a round of interactions and compare with baselineMeasure both the tab and GPU processes
- All three holdAccept as a single-run numberStill compare against the other build, interleaved
- Any one condition failsReject and measure againNumbers are not corrected afterward
- The app does not return to baseline after a roundLook for what accumulatesAbsolute values are not compared
Measuring the two builds in turn within the same window puts noise on both at once Related method
Noise floor first
Round one
Later rounds
Reading
Build A
Build B
1
Build AMeasure A repeatedlySee how widely one build scatters on its own
2
Build AMeasure A once
3
Build BMeasure B right afterNoise lands on both builds
4
Build AMeasure A again
5
Build BMeasure B againRounds follow variation
6
Build AMedian of A
7
Build BMedian of B, compared with AAn improvement only if the gap exceeds the noise floor
Each build is built separately and measured with the same script; the number of rounds follows the noise floor and variation
SEO, AEO and GEO Explained: The Differences and the Order of Work
SEO, AEO and GEO share one body of content and one index; the difference lies at the exit Related method
The website side
- Crawlable HTML
- Direct answers and facts
- Structured data
Crawl
Search and AI engines
- Crawling and indexing
- Retrieval and ranking
Present
What the user sees
- Ranked linksSEO
- Direct answerAEO
- Citation in an AI answerGEO
Schematic flow. Each engine indexes and ranks in its own way.
The order of work: indexing before content, content before markup Related method
Base
Content
Markup
Technical
Content
1
TechnicalIndexingCrawlable HTML, noindex, 404, hreflang
2
ContentMatch the questionTitle, h1, buyer wording
3
ContentAnswer firstOne paragraph per page
4
ContentCitable factsSources, figures, dates
5
TechnicalMarkup and crawlersStructured data, crawler rules
Each step builds on the one before it.
GLB Optimization, Before and After: Download Size vs GPU Memory in Real Projects
Before and after, row by row
| Metric | Before | After | How measured |
|---|---|---|---|
| GPU texture memory: 8 textures reduced from 4096 x 4096 to 2048 x 2048 (WebP quality 80) | about 681MB | about 170MB | Calculated with the formula textures x edge² x 4 bytes x 4/3 (mipmap) |
| Texture download size | 3.38MB | 2.84MB | Texture file size |
| GPU memory: after switching to KTX2 (ETC1S, with mipmaps) | about 170MB | about 30MB | Estimated; these textures are already in the production release. The two steps combined go from about 681MB to about 30MB |
The first meeting takes 15 to 30 minutes and gives a direction, a budget range and a draft timeline.
Polish 3DPolish AINext case: A four-step questionnaire returns a tiered menu, and a 3D scene lays out the whole table
(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
