AI competitions
Public sector AI challenge
We design and run challenges that attack problems in administration, citizen services, and policy using open government data and Korean LLMs. Data sovereignty and security hold, and what comes out can go straight into a real system.
Overview
Open data and Korean LLMs, built to be used
Teams work with high-value datasets released on the Public Data Portal (data.go.kr) and Korean foundation models. The goal is not a leaderboard score but a RAG system, chatbot, or forecasting model that can be dropped into a public service. Because sovereignty and security matter here, Korean models are encouraged and score extra, rather than foreign commercial ones.
Sovereign AI
Why Korean LLMs
The moment sensitive public data leaves for a server abroad, sovereignty and security are no longer yours. The government has backed domestic AI through its sovereign foundation model program, and LG AI Research and Upstage were both selected and carried through to the second round. We use these Korean LLMs as the default engine for our challenges.
LG AI Research · EXAONE
EXAONE is the large-scale model built in-house by the AI research institute backed by 16 LG group companies. Everything from data collection to training was done domestically. It reads Korean accurately and ships with open weights, which makes it a fit for on-premise deployment.
- Built from scratch in Korea
- Tokenizer tuned for Korean
- Open weights, on-premise ready
Upstage · Solar
Solar is the large language model from the company founded by the researchers who led NAVER CLOVA, alongside its Document AI. It performs strongly on Korean and specialist domains, and its document recognition runs at roughly 99% accuracy, which matters when an agency processes documents by the thousand.
- Strong on Korean and specialist domains
- Document AI and OCR
- First generative AI supplier to the Public Procurement Service
Data sovereignty · Korean models and on-premise deployment remove the risk of sensitive data leaving the country at all. Korean performance · Accuracy holds up on Korean-language material such as administrative documents and legislation. Policy fit · It matches the government's domestic-AI-first stance and existing public sector references.
Open Data
How the open data gets used
High-value datasets released by ministries and local governments on data.go.kr become the challenge data.
Collect
Pull administrative, disaster, welfare, and transport data via Open API or file download
Clean
Handle missing values and outliers, then standardize for training and retrieval
Load into the model
Index as RAG documents in a Korean LLM, or use as training data for a forecasting model
Build the service
Turn it into an administrative chatbot, forecast, or automation, and demo it
How We Run
Seven stages, from the brief to what happens after
Define the problem
We surface real pain points in administration, citizen services, and policy, then fix the datasets and the quantitative and qualitative scoring.
Recruit and release data
Datasets and rules go up on the Public Data Portal or a dedicated platform, and teams sign up.
Prep sessions and mentoring
Workshops on Korean LLM APIs, building RAG, and preparing open data.
Qualifying round
Teams advance on leaderboard score or on a written solution proposal.
Finals
Finalists sharpen their models and present the solution with a live demo.
Judging and awards
Scores and presentations are weighed together, and awards from the host agency and the responsible ministry are given out.
Afterward
Winning entries become proofs of concept in real administrative systems or go into public services.
Tracks
Tracks we suggest
We shape the tracks around your agency's data and what you're trying to achieve.
Citizen request triage and reply bot
Route and answer citizen requests around the clock using local government data and Korean LLM RAG.
Document search and summary assistant
Build a knowledge bot that searches and summarizes legislation, notices, and internal rules through RAG.
Disaster and safety forecasting
Build forecasting models for rainfall, flooding, and wildfire on open data.
Finding people welfare misses
Use administrative data to find isolated and vulnerable people early and match them to support.
Joined-up administrative services
Connect scattered procedures for birth, childcare, and housing so they run as one.
Voice phishing and fraud detection
Detect fraud and anomalous patterns in real time from voice and text analysis.
Policy impact analysis
Analyze policy effects and scenarios against public statistics.
Transport and urban big data
Forecast demand and tackle urban problems with land and transport data.
Korean LLM knowledge bot
Build a RAG system on EXAONE or Solar over a ministry's internal knowledge.
Judging
A sample scoring sheet
- Problem solving and accuracy30 pts
Measured performance, or how sound the solution is
- Use of open data, and originality20 pts
- Fit with real administrative work20 pts
- Technical completeness (RAG, model, demo)15 pts
- Korean LLM, sovereignty, and security (bonus)10 pts
- Presentation5 pts
* These weights are a composite of common practice at open data and AI competitions in Korea, and are adjusted for each event.
Impact
What you get out of it
Pain points tested fast
Administrative problems scattered across departments get a proof of concept quickly, from outside minds.
Change citizens actually feel
When winning work goes live, it shows up as shorter processing times and fewer documents to submit.
Better data, because it's used
Real demand for open data feeds back into its quality and into releasing more of it.
AI capability that stays in-house
Your own staff build AI skills, and the domestic AI ecosystem gets a real workload.
We'll run your public sector AI challenge, start to finish
Designing the problem, setting up open data and Korean LLMs, training participants, judging, awards, and the proof of concept afterward.