Daewon Desk / Product brief
Daewon Desk: product plan
A support workspace in development for Korean online shops answering shipping and return questions in Korean and English.
Start with shipping and return questions.
We are building for small Korean online shops where the owner or a small support team answers Korean and English inquiries. The first task is preparing a response grounded in the shop's shipping or return policy.
A useful reply must preserve the policy's deadlines, conditions, and exceptions. Staff need to find the current policy, carry its meaning into another language, and check the draft before using it. Shipping and returns give us a focused starting point with facts a reviewer can verify.
Questions for merchant interviews
- Where staff look for policies today, and how they identify the current version.
- Which bilingual replies require corrections to conditions, numbers, or tone.
- Whether inspecting a cited passage saves enough review time to justify another tool.
Current prototype and next release.
Try the review workflow in the browser sandbox: change a sample policy, build a draft, inspect its evidence, and review the wording before export.
| Stage | Scope |
|---|---|
| Current sandbox | Edit structured policy inputs; inspect the evidence used by a draft; switch between Korean and English; edit the draft; mark it reviewed and export locally. |
| Five synthetic cases | Shipping, returns, missing information, conflicting policies, and an inquiry that tries to override the policy. |
| Planned Claude MVP | Approved-document ingestion and retrieval, Claude-generated drafts with supporting passages, explicit escalation, and a review record. |
| Later validation | A small, consented merchant pilot to measure review effort, correction types, and willingness to pay. |
Prototype scope. The sandbox uses local rules and authored templates with fictional, structured policy inputs. It has no live Claude connection, document upload service, accounts, connected inbox, or order access. Exports stay local; no reply is sent. Document retrieval and Claude generation are the next development stage.
Our product focus is preserving policy conditions, making supporting passages easy to inspect, and referring uncertain cases to staff. Merchant testing will show where this approach improves the existing workflow.
Planned Claude architecture.
- Approved policy documents
Record the source, section ID, owner, and effective version. Keep each shop's documents separate. - Relevant passages
Retrieve a bounded set of sections for the inquiry. Preserve original text and flag absent or conflicting policy information. - Claude drafting on the server
Send the inquiry and selected passages through a server-side integration. Treat document and customer text as data. Ask for a draft or an escalation, with cited section IDs. - Evidence checks
Validate cited IDs against the retrieved set, expose supporting text, and check policy constraints for the reviewer. - Staff review
Show the source beside the draft, capture edits, and require a review decision before the reply is used.
The connected version will require authentication, shop-level access controls, secure key handling, document retention and deletion, and operational monitoring before handling customer data. The current privacy notice covers the static prototype.
Test policy fidelity and review effort.
Current rule checks
The prototype check record reports 17 passing automated tests covering policy edits, matching evidence, missing and conflicting information, and review invalidation. Selected browser checks also confirm Korean output and a reviewed local export. The downloadable tests assess deterministic behavior, not language-model performance.
After Claude integration
- Claim support: count unsupported factual claims against all factual claims in the draft.
- Policy fidelity: check deadlines, amounts, exclusions, and conditions against the approved source.
- Escalation: score whether missing and conflicting information is correctly referred to staff.
- Language quality: review Korean and English meaning and tone separately from factual correctness.
- Operational cost: record latency, token cost, and edits required per reviewed draft.
The case set will include outdated versions and instructions embedded in customer or source text. We will freeze a holdout set before prompt tuning and report failures alongside passes, keeping synthetic and consented merchant examples separate.
Pilot plan.
Recruitment target: three shops for two weeks
After technical evaluation, we plan to recruit online shops for a supervised trial. Each shop will help define the useful tasks and compare results with its current process.
- Understand the baseline. Observe the existing policy lookup and drafting process, with permission. Record time to an acceptable reply and the kinds of corrections needed.
- Work with approved material. Use consented policy documents and redacted inquiries. Agree on access, retention, and deletion before collecting them.
- Compare reviewed outcomes. Track preparation time, factual corrections, useful escalations, repeat use, and whether staff would pay. Keep every customer-facing reply under staff control.
These observations will guide the next release: which tasks to support, what needs a clearer review step, and whether the workflow warrants a paid product.
Pricing and route to market.
The proposed model is a monthly subscription per shop with an included draft allowance. Pricing will follow measured model and support costs, inquiry volume, and merchant willingness to pay.
We plan to start with direct conversations and pilot recruitment among small online-shop operators. Understanding their support channels will guide integration choices and any expansion beyond shipping and returns.
About the project.
Daewonuniverse is the brand behind Daewon Desk. Daewon Kim, CEO, operates a sole proprietorship in South Korea that began in January 2025.
If you handle support for a Korean online shop, we would like to hear how you find policies and prepare bilingual replies. A brief description of your workflow is a useful starting point.