Welfare Compass v2 — Production Service
Team of 3 · Full-Stack Rebuild · Jan–Mar 2026
Problem
The hackathon prototype had no persistent state and a simplistic retrieval pipeline — couldn't scale for real users with diverse eligibility conditions.
Approach
- Django + Next.js full-stack architecture with persistent user profiles and eligibility engine
- LangGraph ReAct Agent replacing sequential RAG with autonomous reasoning + tool use
- Hybrid search pipeline: BM25 + Dense (BGE-m3) + Cross-encoder reranking
Outcome
Architecture designed and core modules under active development. Production deployment planned for March 2026.
Tech Stack
Business Impact: Evolving a hackathon prototype into a production-grade service — persistent conversation flow, eligibility-aware recommendations, and scalable architecture for real-world deployment.
Welfare Compass v1 — Hackathon Prototype
Seoul AI Hackathon · Top 20 / 181 teams · 2025
Problem
300+ welfare policies scattered across documents — citizens struggle to determine eligibility, especially those with low digital literacy.
Approach
- RAG pipeline using LangChain + FAISS for policy document retrieval
- Natural language query processing to extract relevant policies and eligibility criteria
- Streamlit-based conversational interface for accessible interaction
Outcome
Demonstrated effective retrieval for complex policy queries, advancing to hackathon finals.
Tech Stack
Business Impact: Reduces citizen service center workload by automating policy matching — a scalable RAG solution applicable to any document-heavy consultation workflow.
CatchUp v2 — Study Material Structuring Pipeline
Solo Project · VLM/LLM Pipeline · Mar 2026
Problem
Dropping a document into GPT gives a summary — but when you have a PDF, a notebook, and screenshots covering related concepts, nothing connects the knowledge across them.
Approach
- Multi-format parsing pipeline: PDF (DoclingLoader), Jupyter notebooks (nbformat), images (VLM 5-class classification)
- 12 VLM models benchmarked across 8 axes — including reasoning vs. non-reasoning and generational gap analysis
- LLM-powered note generation with versioned prompts (v1.0→v1.4) and concept extraction with cross-document backlinking
Outcome
Core parsing + VLM client operational with 10-model unified API, cost tracking, and structured JSON output. Experiment and RAG pipeline under active development.
Tech Stack
Business Impact: Transforms scattered study materials into a searchable, concept-linked knowledge base — demonstrating multi-modal document AI pipeline design with systematic model evaluation across 12 VLMs.
Drama Color Pattern Analysis
Published · Journal of the Korean Data Analysis Society · Feb 2025
Problem
Color's impact on viewer emotional engagement was based on subjective judgment — no quantitative analysis existed for cross-cultural comparison.
Approach
- ConvNeXtLarge for scene clustering across drama episodes
- HSV/LAB color metrics comparing Korean vs Spanish versions of Money Heist
- User survey (n=86) for preference validation
Outcome
Korean color grading preferred 1.5× overall across the full sample (p<.05). First author publication.
Tech Stack
Business Impact: Data-driven framework for visual content optimization — applicable to automated color grading pipelines, content recommendation, and A/B testing visual assets at scale.
Welding Defect Detection
Presented at IPIU 2024 · First Author
Problem
No real defect images available in manufacturing environments due to security and cost constraints — severe training data shortage.
Approach
- CycleGAN for domain adaptation (X-ray radiographs → inspection images)
- Preprocessing pipeline (contrast enhancement, wavelet filtering) to suppress artifacts
- YOLOv5 with two-stage transfer learning (COCO → synthetic defects)
Outcome
87.5% mAP achieved on synthetic-to-real defect detection. First author conference paper.
Tech Stack
Business Impact: Eliminates dependency on restricted real-world defect data — enabling automated quality inspection deployment in manufacturing environments where labeled data is unavailable.