The Geometry Measures, the AI Teaches: Building an Art Studio Tutor with ADK, Vertex AI & OpenCV

A 9-year-old daughter's growing interest in drawing led to the creation of Atelier, an AI-powered art studio tutor. The project was born during the Devpost All Things Agentic Hackathon. Atelier bridges a common issue in remote art education: students cannot see their own angular deviations in drawings.

Atelier is an agentic AI studio master that decouples deterministic geometric calculation from pedagogical critique. It uses OpenCV for geometry measures and Gemini Flash on Vertex AI for empathetic critique. Atelier's architecture consists of two planes: Plane A, which contains deterministic findings derived from OpenCV metrics, and Plane B, where the LLM evaluates the drawing qualitatively.

The two-plane critique model ensures that any numerical measurement mentioned in the critique is strictly derived from OpenCV metrics. If a mismatch is found, the response is rejected and regenerated. Atelier employs four collaborative verbs: ASK, GUIDE, CAPTURE, and ADAPT. ASK is used to calibrate feedback depth, GUIDE prescribes specific drills, CAPTURE collects explicit feedback, and ADAPT dynamically adapts to the student's learning profile.

Atelier's asynchronous pipeline starts when students or parents upload their sketchbook photos to a private Google Cloud Storage bucket. Eventarc then triggers Cloud Run to perform geometry calculations and critiques in the background. Weekly aggregations calculate the error reduction percentage, and a 3-day practice plan is prescribed for the upcoming week. The system is deployed on Google Cloud Run with .NET 10 and Python FastAPI microservices.