Skip to main content

Research

From venue delivery to Xiyu: product practice through an FDE lens

Why can a completed order still have unused services? A delivery problem leads into business modeling, market judgment, and industry AI.

8 min read
On this page

An order can say “completed” while the player still has services left to use.

Both can be true. A player may buy several instant-photo or joint-photo services and redeem them at different times. The person who sold the service may not be the person who performs it. At settlement, the venue must still explain who did what, which rule applies, and where the original record can be found.

Many problems we encountered in a custom venue system lived inside apparently contradictory states like these. Opening a page and accepting a payment were only the beginning. Revisions after the summer delivery, followed by product observations and merchant material in September and October, made one requirement increasingly clear: an industry product needs to explain how a piece of work actually gets finished.

We later began reviewing this experience through the lens of FDE—Forward Deployed Engineering. The useful idea is to learn the business through delivery, expose the implementation to actual use, and bring recurring problems back into the product. Custom work supplied that understanding; Xiyu is our direction for extracting common foundations and exploring industry AI.

“Completed” needs more than one record

Start with a small example.

Suppose a player buys three services, uses two, and saves one for later. These quantities only illustrate the model. An “already used” switch on the order loses part of the truth whichever way it is set.

Duplicating the order into three rows is not enough either. All three services came from one purchase, and refunds must refer back to its line items. Different staff may perform each service, so the eventual contribution cannot always be fixed at purchase time.

We therefore distinguish several records. An order records the purchase; an entitlement records what remains available; a redemption records what was actually delivered; a settlement records the basis for confirmed compensation. They refer to one another rather than replacing one another with a single status.

Domain relationships from order lines to entitlements, redemptions, and settlement details, with staff and rule references
Figure 1. Conceptual relationships derived from delivery problems. 1→N means one record may have multiple downstream records; dashed lines indicate contextual and rule references. This is a business model, not Xiyu’s current database schema or a list of available features.

The separation has a practical purpose. Remaining entitlements answer the player’s question about what is left. Redemption records tell reception whether a service was performed. Settlement details tell finance where a figure came from. Different people need different answers about the same transaction.

Historical operating material put some of these rules into issuance settings, validity periods, and staff binding. They look like configuration fields, but they determine whether a service can be used accurately later.

Entitlement configuration in a historical custom project
Historical project material, in Chinese: issuance, validity, and staff binding. It explains the rules of that project, not today’s Xiyu interface; the displayed duration is not a current product promise.

Exceptions test the model

A design can look complete along the normal path. Staff substitutions, partial refunds, and changes after settlement reveal what it missed.

Return to the three-service example. Two services have been delivered, and the player requests a refund for the remaining one. The order now has a refund, and one entitlement becomes invalid. The two completed services still exist. A refund must neither erase them nor make the same entitlement usable again.

If a confirmed settlement then needs adjustment, the earlier result must remain explainable, and the adjustment must have a reason. The reversal method and approval role depend on the business agreement. Simply overwriting the old number removes the basis for the next reconciliation.

Swimlane sequence for partial use and a refund of the remaining service across player, reception, records, and finance
Figure 2. An illustrative purchase of three services, with two delivered and the remaining one refunded. Solid paths show normal fulfilment; the orange branch shows exception handling. Refund scope, approval, and settlement adjustment require agreement with the venue. No actual transaction is represented.

This is where permissions enter business design. Who confirms delivery, initiates a refund, or changes attribution cannot be answered solely by saying that an administrator can do everything. The operator, time, original basis, and reason for adjustment all belong to the work.

It changed our requirements discussions. We began asking less about another column in a list and more about what the next colleague would still need to continue the task.

The deeper the delivery, the more carefully we must reassess the market

One project can provide depth. It does not establish that every venue needs the same product.

In September, we began scoping an independent product and observed public pages from two performance-venue mini-programs: how bookings opened, how roles were chosen, and how services appeared. October conversations and operating material added another view. Staff submitted forms, spreadsheets calculated figures, finance checked them, and vacancies continued circulating through chat groups and posters.

Each source served a different purpose. Delivery records revealed changes in use. Public pages exposed different booking arrangements. Operating material showed collaboration beyond the interface. Together, they still do not establish market share, willingness to pay, or growth effects.

ObservationEffect on product judgmentWhat still needs validation
Popular sessions open for booking togetherExplain opening times, remaining roles, and confirmation rulesHow reservations, changes, and exceptions are handled beyond the page
Unfilled sessions recruit continuouslyUse session and vacancy information as a common source for recruitmentHow much information still needs manual updating after a change
Staff report, spreadsheets calculate, finance checksPreserve source details and accommodate established workWhether a new tool reduces reconciliation and repeated entry

This clarified the product boundary. Sessions, orders, services, and people are candidates for shared foundations. Session fees, contribution rules, membership thresholds, and brand expression need room for venue-specific configuration.

Market assessment also returns to switching costs. Spreadsheets and group chats already constitute a working system. A replacement must make an important task more reliable or easier before asking a team to change its habits. A longer feature list cannot establish that value by itself.

Once AI joins the work, who is responsible for the result?

These business relationships give AI a concrete task.

“Help recruit players for Saturday evening” requires at least the venue, game, time, available places, and confirmed instructions. Asking a model to infer them from conversational context can put outdated availability or incorrect rules into a polished poster.

The direction we want to explore starts with business records the user is allowed to read. AI helps organize the expression; a responsible person checks the content. When a business action is needed, it goes through a controlled tool, which returns its actual result. Generating content and changing business records each need explicit responsibilities and checks.

Reference architecture connecting scoped business sources to model and tool orchestration, with separate paths for answers, drafts, and approved actions
Figure 3. A reference architecture for future Xiyu integration. It separates source information, permission scope, models and tools, human confirmation, and result records. The entire diagram is a design direction, not a deployed architecture or production acceptance report.

The first step can be narrow: answer an operating question and let a person find the statistical scope and source. Drafts for copy, posters, or tasks can follow. Confirmed actions come later. Reading, drafting, and acting are different tasks; errors, rework, and usability should be observed separately.

For a service-record query, checks might include missed items and consistent calculation rules. For recruitment content, they might include accurate time, availability, and brand information, plus the editing needed to reach a usable version. Whether the capability deserves further investment depends on both the model’s output and the whole task.

This explains the connection to historical systems delivery. It taught us where information lives, who can confirm it, and which changes affect someone else. Model capabilities still need integration and validation. The business foundation provides specific situations in which to test them.

What FDE brings back to the product goes beyond code

In its October introduction to FDE, OpenAI discusses working with customers on concrete problems and bringing the learning back into the product. That perspective echoes our delivery experience: a seemingly local revision can force a team to rethink the work itself. Read the source.

We want to retain more accurate business objects, processes that accommodate exceptions, and judgments about which differences deserve to remain. Less visible delivery work matters too: confirming rules, providing operating instructions, establishing feedback routes, and knowing who takes over when something fails.

This retrospective uses FDE as a working lens. It does not claim multiple production AI deployments or relabel historical custom work as commercial results for the standard Xiyu product.

The next validation can still begin with a small question. When availability changes, a staff member is replaced, or one service is cancelled, can the people involved continue correctly? After AI joins, is the task easier to explain, or does it create another step someone must repair?

These are the questions Xiyu must keep answering when the venue opens again.

Based on delivery and research material through the end of October. Numerical flows are illustrative. The cover and diagrams were produced in November.

View all