1
GuestFour-step questionnaireAnswered with cards and buttons
(Case study) Catering・Polish AI / Polish 3D
A four-step questionnaire returns a tiered menu, and a 3D scene lays out the whole table
Guests booking wedding catering need to confirm the menu and also see how the table will look. Guests describe the event in four steps, receive a tiered menu, and move straight into a 3D scene to lay out the whole table.
The same approach suits caterers, banquet venues, wedding planners, tableware and floral rental, and any service that is understood only once the parts are put together.
7catering services
6preset packages
4question steps
(Challenge)
Catering guests need to confirm both the menu and the look of the table
Catering guests buy a picture of the whole meal: the food and the look of the table on the day. There are seven services, from Western-style weddings, round-table banquets and first-birthday parties to individual boxed meals, each with its own specifications. Date, headcount, venue and dietary restrictions all change the menu and the price.
When each inquiry is handled by hand, one question and one menu at a time, proposal speed affects how many orders can be taken and how long guests spend comparing prices.
(Approach)
AI writes the menu, and prices always follow the specification
The biggest risk of putting AI in front of a guest is that it quotes prices on its own. The work is therefore split in two. Pricing method, per-person price and add-on services follow the back-office specification, and the number of dishes is trimmed to the specification limit. AI handles only package names, menus and descriptions, and the back end overwrites any price that differs from the specification.
The dishes AI draws on are filtered first: the season is inferred from the date, indoor or outdoor from the venue, and vegetarian needs from the description. A pool of candidate dishes is then picked from the menu library and handed to AI.
Proposals step up by specification from classic to upgraded to chef's selection, and the system also gives AI past cases with a similar service, price range and headcount for reference.
Guest
- Four-step questionnaire
- One-line descriptionFor example, many elders or a vegetarian guest
NeedsTiered packages
Development team's service
- Candidate dish filterBy season, indoor or outdoor, and vegetarian needs
- Spec overridePricing method, price per person, add-ons and the cap on courses
Candidate dishesNames and descriptions
Model provider
- Language modelPackage names, menus and descriptions
Not left to the modelPrices, pricing method and the cap on courses.
Scope: proposal generation.
The menu is written by AI, and the price is set by the back-office specification.
Read the full article RAG for Enterprise Knowledge Q&A: A Verbatim Verification Workflow(Key design)
Three key designs
Four-step questionnaire
Answers are mostly cards and buttons. Available dates come from the system in real time, and a refresh halfway through loses nothing. No registration is needed, and an email check after submitting saves the proposal. The last step takes a single sentence, such as many elderly guests or some vegetarians, and AI adjusts the dishes, reserves vegetarian options and avoids the dietary restrictions mentioned.
3D table
Guests who chose a Western-style wedding, round-table banquet or buffet confirm a package and carry the order straight into the 3D table. Every table type has built-in seat anchors, so chairs, plates, glassware, napkins and menu cards snap into place. Swapping a square table for a round one rearranges the whole setting automatically.
Preset packages
Six preset packages apply in one tap, changing furniture, tableware and flowers together. Guests who prefer a custom setting can replace items one by one across ten categories. When the setting is done, the selections are saved to the same order, and the team sees the chosen package and items in the back office.
Desktop
Six preset packages, the whole table swapped at once
Six preset packages, three for square tables and three for round ones, are listed under the 3D table. Selecting one rearranges the whole table, including the table type, chairs, tableware and runner. Switching to customization allows each category to be replaced item by item.
Phone
On a phone, the whole table also changes in one tap
The 3D table on a phone keeps the same presets and the customization tab. One finger rotates the view, and selecting a package rearranges the whole table.
For services that only make sense once assembled, the first meeting can break such a service into a questionnaire and rules.
(Result)
Two connected websites and an AI that can be tuned
There are now two linked websites and an AI that the team can tune itself. Prompts such as the chef's voice and naming rules are managed in sections in the back office, and every edit keeps a version that can be rolled back. Items and packages are added or removed there too, so a new plate needs no code change.
3D models are compressed and step down automatically on phones, so guests can lay out a table on a phone. Introductions and Q&A are written into the page source, where search and AI engines can read the content.
Questionnaire
Proposal
Setup
Confirm
Guest
System
Team
2
SystemTiered packages producedFrom Classic and Upgrade to Chef's Selection
3
GuestEnter the 3D tableApply a preset or replace item by item
4
SystemSelections saved to the order
5
TeamTeam views packages and items in the back office
The 3D table suits Western weddings, round-table banquets and buffets.
Services and FAQs come from the back-office database, while the text stays in the page source.
- Content sourceStatic files, or a back-office database
- Update frequencyFixed, or frequently updated
- When 3D fails to loadWhether users can still read the text
- Content is fixed and the number of routes is limitedBuild-time prerenderingPOLISH's own site: no framework change, one build step renders each route to static HTML
- Content comes from a database and changes oftenServer rendering with incremental static regenerationService cards and FAQ are served from the back-office database
- 3D is the main interface and loading can failServer text layer, removed once 3D takes overA public-sector site: text shows only if 3D fails to load or times out
- Text is placed only for crawlers and users cannot see itNot adoptedContent must match what users see
(Field notes)
Methods and measurements, open each one
How an Online Catering Proposal System Works: A Four-Step Form, Spec-Based Pricing, and a 3D Table
Prices, dish counts, and add-ons come from the admin specs; the model returns only names, menus, and descriptionsAfter the model responds, every item is checked against the specs, and the admin values decide amounts and dish countsMaking 3D Websites Readable to Search and AI Crawlers: Prerendering, Bot Policy, and Structured Data
Prerendering raised the raw HTML of POLISH's home page from 730 bytes to 39,516 bytes with no visible change Related method, figures from another project
Raw HTML size of the home page (bytes)h1 count in the raw HTML
- 730 bytesBefore prerendering (single-page app, empty container)Before0
- 39,516 bytesAfter prerendering (static HTML per route at build time)Adopted1
POLISH's previous home page. The h1 counts follow the described prerendering approach.
The three crawler types read the same HTML for different purposes, and robots.txt covers each differently Related method
Crawlers
- Search indexing crawlersOAI-SearchBot, Claude-SearchBot, PerplexityBot: fetch text for search and assistant results
- Training crawlersGPTBot, ClaudeBot; Google-Extended is a token with no user agent of its own
- User-triggered crawlersChatGPT-User, Claude-User, Perplexity-User: fetch one page only when a user asks
Requests and user agentHTML text
Website (set up by the development team)
- Prerendered HTMLThe h1, direct answer, and body text are in the source
- robots.txt groupsSearch and assistants open, training by business decision
- User agent check at the serverRefuses requests that ignore robots.txt
robots.txt is not access controlRFC 9309 states that its rules are not access authorization; for user-triggered fetches, the official documents say the rules may not apply.
Compiled from the official documentation of OpenAI, Anthropic, Perplexity, and Google, checked on 2026-11-05
Estimating the Cost of an AI Project: Why Estimates Miss by Several Times
With one verification agent per finding, the agent count grows with the findings to 136; batched verification brings it down to about 7 Related method, figures from another project
- One verification agent per finding136 agentsAgent count scales with findings
- Agents removed by mergingCost at about 5% of the originalOne agent verifies 15 to 20 findings
- Batched verificationAbout 7 agentsCeiling fixed before dispatch: findings times verifiers
Audit workflow. Costs are converted at list prices, not invoiced.
Turning thinking off cuts writing time to 6 to 17 seconds, but heading success drops from 11/11 to 2/13 Related method, figures from another project
Writing timeHeading success rate
- 15 to 75 sThinking on, single passKept11/11
- 6 to 17 sThinking off, single passNot adopted2/13
Test site, real model, 13 questions. Sentences removed by the validator were 3 times as many with thinking off.
The most expensive model handles only the hard cases flagged in sampling, not every document in the corpus Related method, figures from another project
- Tag the full corpusAbout 1,500 documentsOne subtask per document, tagged by Haiku
- Documents not sampledValidator onlyAccepted at exit code 0
- Risk-based samplingAbout 160 documentsAbout 10%, reviewed by Sonnet
- Escalate to Opus9 subtasksOnly flagged hard cases and anything the validator rejects
Subtasks: about 1,600 for Haiku, about 50 for Sonnet, 9 for Opus. Document and subtask counts are approximate.
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 answerSEO, 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 exitThe first meeting takes 15 to 30 minutes and gives a direction, a budget range and a draft timeline.
Polish AIPolish 3DNext case: An annual cross-ministry outdoor education program, linking venues across Taiwan with a no-download web game
(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
