Inspiration
SettleK is for foreign residents living in Korea. People have different residence statuses, housing arrangements, work situations and plans, yet finding the official information relevant to their own circumstances can be difficult. Conversations with an American friend and observations of international students helped spark the idea, but the audience is foreign residents generally.
I wanted to connect understanding a situation with actually preparing the next step, and to help people keep their progress when their circumstances change.
What it does
SettleK turns a structured situation and an optional consented note into a personalized preparation plan. Real AI extracts a summary and supported facts; follow-up questions clarify missing or conflicting answers. Reviewed rules connect those answers to preparation candidates and scoped official references.
The central demonstration follows a fictional D-10-1 resident who is looking for work and then receives a job offer. Updating the employment answer shows two preparation candidates added, one removed, and two carried forward. The lease agreement keeps its ready state, so the person can see what changed without losing work already done.
Document cards provide preparation guidance, official links and bilingual request text. Users can track documents they hold, confirm reuse, save and restore account progress, export a backup, and create reminders for dates they choose. Calendar export, history, Korean/English interfaces and mobile installation guidance complete the nine-screen flow.
SettleK supports preparation and official verification. It does not decide visa eligibility, grant immigration permission or submit an application.
How I built it
I used the Devpost Learn Skill Pack from 1-start through 6-ship in a new project folder, with Codex as the coding assistant. The repository includes scope.md, prd.md, spec.md, the build checklist, an app map and English planning counterparts.
The app uses React, TypeScript, the Sites Vinext starter, Tailwind CSS and Cloudflare D1. Server-side calls to Nebius Token Factory use NVIDIA Nemotron 3 Super for bounded fact extraction. Reviewed rules and official-source references determine the preparation candidates. API keys remain on the server, and account records are separated by authenticated user identity.
Challenges and what I learned
A key challenge was making changes understandable while preserving useful progress. The before/after comparison and separate ownership, preparation and reuse states address that problem.
Testing also made the distinction between AI analysis and an answer-only update important: a new AI request can read a consented note, while changing selected answers updates the plan without another model call. The interface labels these paths separately. Recording the real workflow helped refine the instructions around questions, document preparation and saving.
What is working
The project has 82 passing automated tests, plus recorded live model and hosted account checks. The creator's demonstration shows real AI output, the changing preparation list, official links, retained document readiness, calendar import and successful account saving. Automated provider tests use mocks; the verification document identifies the separate real requests. External foreign-resident usability testing is a next step.
What's next
Test the flow with foreign residents, improve wording from their feedback, and expand the reviewed residence-status guidance with clearly scoped official sources.
Prior work disclosure
The SettleK name, concept, design direction and user journey predate this entry. The application code was newly written for this Skill Pack build; the existing SettleK implementation was not copied. The Sites Vinext starter and dependencies are retained foundation code. The public repository is a source snapshot, with its provenance explained in the documentation.
Log in or sign up for Devpost to join the conversation.