Internal Open Call and Training Operations System

Overview

Combined a manual workflow into a single web service. In that workflow, quarterly open calls were received by official memo and email, first-round evaluation and score collection by business unit were handled in Excel, and results were confirmed after a second-round evaluation. The goal was a general-purpose system that can handle the application questions and evaluation procedures that differ from one call to the next.

Handled everything alone in a 5-week individual project: organizing requirements, the React screens, the FastAPI server, testing and in-house deployment.

Key features

• Application forms for each call: Set up the application questions to fit each open call or training program, and receive applications
• Stage-by-stage evaluation: Connects reviewer assignment, multi-stage evaluation, and result tallying and confirmation
• AI evaluation assistance: The in-house RAGaaS provides application summaries, question answering, and reference scores and source evidence for each question

Implementation and improvements

1. Separating the common work flow from the differences between calls

   Organized the common flow as announcement → application → evaluation → result confirmation. Application questions are defined with a JSON schema, and evaluation stages with order and type settings, so that a new call can be opened without code changes.

2. Connecting the evaluation procedure to roles and stages

   Connected role-based permissions, reviewer assignment, multi-stage evaluation and result tallying. The server checks roles and assignments, so a reviewer who is not assigned cannot access an application.

3. Building AI features that help evaluators decide

   Developed a React-based screen and a FastAPI-based backend, and connected them to the in-house RAGaaS (an internal system that gives AI answers based on company documents). It provides application summaries, question answering, and reference scores and source evidence for each question, while only human scores go into the average and the final confirmation. Evaluators and administrators check the source text and the AI reference information before confirming the final result.

4. Improving the first screen for each role based on feedback

   At first, applicants, evaluators and administrators started from similar screens. In a mentor review, the feedback was that it was unclear where each of them should start their work, so the first screen was split by role and the menu names and evaluation procedure were organized. Afterward, the complexity of the score entry screen and the mixed use of terms were also fixed.

Tech stack

Results and current status

Integrated everything from application through evaluation to result confirmation into one web service and deployed it in-house. The responsible department is now improving it on that basis and using it in real work. Business rules and permissions were verified with 205 pytest tests and 130+ Vitest tests, and the flow from submitting an application through evaluation to adjusting results was checked with Playwright.

Real screens and internal materials are not shown, in line with the company's security policy.