Air Force AI Service Development

Overview

Handled feature definition, UX design and frontend development for several internal AI services during Air Force service. This page covers several services together. The process of improving chat screen performance is written up separately in AiRWARDS 3.0.

On AiRWARDS, designed the chat interface and the state management structure. Development contribution by page was chat 80%, document summarization 50% and ASR 50%. On the real-time AI translation system and the AI-based National Assembly work management system, took part in development covering 100% of the frontend and 30% of the backend.

Key features

• AiRWARDS 3.0: Conversational AI, document summarization and speech recognition used on the defense network (a military internal network separated from the outside)
• Real-time AI translation system: Shows continuously arriving utterances and their translation results
• AI-based National Assembly work management system: Visualizes keyword relationships in administrative work data in 3D
• Runway crack detection system: Checking detection results on high-resolution images, with zoom in and out
• Air Force data hub: An internal data integration platform that supports SSO (one login for access to multiple systems) and standard login

Implementation and improvements

1. Handling continuously arriving AI responses on screen

   Improved the streaming rendering (showing an answer on screen as it is being generated) and the state management structure of the AiRWARDS chat screen. The specific problem and how it was solved are written up in AiRWARDS 3.0.

   On the real-time AI translation system, handled the frontend alone and designed the state management logic that reflects utterances and translation results on screen. Used Tailwind-based design tokens (base values for color and typography) so that theme colors can be changed on user request.

2. Visualization and authentication built around the work data

   Handled the frontend of the AI-based National Assembly work management system alone and built a 3D word cloud with Three.js. On the runway crack detection system, improved rendering and zoom for images of 7,000×7,000 or larger, and also took part in developing a data preprocessing pipeline to speed up the detection model.

   On the Air Force data hub, designed an authentication structure that supports both SSO and standard login, and built an authentication flow that reflects internal security policy.

3. Proposing React and a shared component registry

   Vue 2, in use since 2019, had few experienced people among the enlisted developers, who change every one and a half years, and it depended on archived libraries. Proposed developing the new systems in React instead of Vue 2. Common components copied from project to project had inconsistent props and UI, so built a shadcn custom registry on the internal network, where centrally managed base components are pulled into each project and customized.

4. AI coding environment on a closed network, and usage analytics

   The network was closed, so external AI coding tools could not be used. Got Continue, Roo Code and OpenCode working with a lightweight model on the internal network by resolving grep blocking, external API call errors and Windows freezing, and wrote shared guideline documents. Also introduced Umami (a web analytics tool), which set up an operational base for checking access to key screens and how features are used.

Tech stack

Results and current status

Worked on improving the chat screen of AiRWARDS, which is used by 5,000+ monthly active users (MAU) on the military intranet, and took part in developing several internal services for translation, administrative work and data use. After AiRWARDS 3.0, built the foundation for 10 or more new React systems and accumulated about 50 shared components. Alongside service development, built shared UI distribution, an AI development assistance tool environment and a user analytics environment.

Letters of confirmation