ArtPrep: AI Fine Arts Exam Gallery
AI artwork diagnosis and an admitted-works gallery for Korean art-school applicants. • Students' feedback on their own drawings depended on a tutor's availability. I built the flow where an uploaded drawing returns a 5-axis diagnosis report via Amazon Bedrock (Claude) and academy recommendations. The diagnosis report shipped. • An admitted-works gallery cannot stay current on manual uploads. I built a crawling and AI-tagging pipeline over 49 sources that runs nightly. 46 of 50 golden-set images pass. • Artwork retrieval and the knowledge graph needed research beyond product work. I defined the problem and supplied data for joint research with BOAZ (student data science society), documented in a joint technical paper. • The team also joined KAIST OverEdge. Tech: React · Vite · Tailwind CSS · Supabase · Amazon Bedrock · Vercel
Metrics
- Diagnosis axes: 5
- Golden-set tagging: 46 / 50
- Crawl sources: 49
- Joint research: 1 paper
Highlights
- 5-axis diagnosis as fixed JSON
- A gallery that refreshes itself
- Research, not just product
- Built so the team can continue
- Represented in KAIST OverEdge
Story
Situation
Korean art-school admissions prep is expensive and fragmented. Students search across sites for reference works, and feedback on their own drawings depends on a tutor's availability. No platform combined a searchable gallery of admitted works with structured AI feedback.
Task
Ship a product where a student uploads a drawing and receives a structured diagnosis and a recommended next step. Keep the admitted-works gallery current without manual uploads.
Action
Own product, engineering, and deployment as CTO in a 3-person team. Built the front end on React, Vite, Tailwind, and Supabase. Diagnosis runs on Amazon Bedrock (Claude) with forced tool use, so the model returns a fixed JSON: 5 axes and academy picks. Built a crawling and AI-tagging pipeline over 49 sources that runs nightly; tagging quality is checked against a 50-image golden set with human review. Defined the research problem and supplied data for a joint study with BOAZ (student data science society) on artwork retrieval and a self-improving knowledge graph, documented in a technical paper. Dev logs and bug reports are maintained in Notion so the team can continue without me.
Result
Diagnosis report shipped: summary cards, 5-axis radar, academy recommendations. Golden-set tagging at 46 of 50. Crawl runs automatically every night. ArtPrep was represented in KAIST OverEdge by a team member. Roadmap through late October 2026: deeper analysis, recommendations, then discovery of works, academies, and instructors.