VENTURE / 셀릿(잇다) · SELIT
I co-founded a video production operations platform. As CTO, I designed the policies and UX connecting intake, production, review and settlement, and built most of the frontend and backend with AI development tools.
PRODUCT OWNERSHIP / OPERATING SCALE
INTAKE / BEFORE & AFTER
These screens compare a product-tag selector with a step-by-step intake flow. The revised screen shows a brand-name question, not AI product classification. The captures use different sample requests and illustrate the interface approaches rather than the same request before and after.
Earlier component: chore/staging-snapshot-before-main-sync-20260811 · e1e8191. Conversational UI: 60f5259. Both captures use sample data; they are component demonstrations, not historical production screenshots or evidence of AI classification accuracy.
The August branch also contains conversational intake code. Its date establishes that this tag selector existed in the repository; it does not establish the production rollout sequence.
01 / DECISIONS
One person handled support, so capacity had to be managed. Direct support work showed me that clients struggled with our product categories. I prioritized changing intake to reduce the need for repeated explanations.
Conversational intake checks for required production information and asks follow-up questions. This shifts responsibility to the product: clients need a way to correct automatic answers and continue when extraction fails.
I gave planning, filming and editing their own owners and status. Eligible freelancers accept each task; approvals and submitted materials trigger the next handoff.
More roles require consistent materials and handoff conditions. Approval checks and operational handling for unassigned work remain necessary.
The data model separates owners, status and candidates. Implementation includes acceptance checks, next-role selection after planning approval and editor notifications after footage submission.
I developed notifications and a scoring system, offering faster handling and price benefits to clients who sent revision requests promptly. This client incentive policy did not penalize low scores.
The operational target was feedback within three hours, to reduce freelancers’ waiting time. The response-target achievement rate and the cost of the incentives have not been established in this case.
Using SQL, I queried and aggregated the time from request submission to first draft, and from feedback receipt to implementation of requested changes. Comparing the two helped me prioritize which stages to assess for AI adoption.
This analysis informed AI adoption priorities; it does not represent measured time savings after adoption.
SELIT / COLLABORATION
I turned the review criteria shared by a co-founder into a product feature.
01 / PRODUCTION CRITERIA
Co-founder · Video production experience
Shared the checks clients care about: spelling, scene order and whether requested changes are reflected.
02 / PRODUCT IMPLEMENTATION
Dohyeon Kim · Product planning and implementation
I designed and implemented an AI feature to check these items before review.
03 / CURRENT STAGE
AI pre-review · In testing
The aim is to catch basic omissions early and reduce avoidable revision requests and repeated reviews.
A reduction in review rounds or an improvement in checking accuracy has not yet been established.
02 / CONNECTED WORKFLOW
Intake connects to role-specific acceptance, handoffs, files, notifications and settlement in the operating platform.
Choose the production scope in language clients understand.
Captured locally on Sep 28, 2026 from current source components. The demo shell, names, prices, deadlines and feedback are fictional. Saving, sending and uploads are disabled. Select a screen to enlarge it.
The admin library brings videos and image carousels together with project context, tags and visibility. The client portal then makes that work browsable.
Reading the Korean interface: each card is a completed video or image carousel. Project details and tags preserve its context; visibility controls determine what appears in the client portal. The media shown here was created by production teams. I built the platform that organizes it.
Project, content type and tags make the output reusable as a reference.
Visibility and active status separate the library from the client-facing selection.
Thumbnail and short-preview fields are separate from full playback in the product code.
Shown work was produced by teams through SELIT. My contribution was the platform’s product structure, role-based UX and implementation.
03 / EVIDENCE & LEARNING
The operating count above comes from a supplied dashboard capture, without a database audit. The original capture is not published. The founder recalls the intake change on August 12, 2026; the rollout sequence was not verified against deployment records.
As supporting context only, I retrospectively estimate daily KakaoTalk support at about four hours before and one hour after the change. This was not measured from time logs. The comparison period, observed days, request volume and follow-up rate are unavailable, so the estimate is not presented as a core outcome.
Compare support time per submitted request alongside total workload, and check how often intake needs follow-up or correction. This would help distinguish less support work from simply fewer requests.
An interactive storybook designed around hand-gesture rehabilitation activities and attention. I focused on comfortable participation for children and understandable results for caregivers.