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

The Geometry Measures, the AI Teaches: Building an Art Studio Tutor with ADK, Vertex AI & OpenCV This project was created for the Devpost All Things Agentic Hackathon. When my 9-year-old daughter began practicing drawing cubes and perspective boxes in her sketchbook, I noticed a universal paradox in remote art education: Students can't see their own angular deviations (a $4^\circ$ or $7^\circ$ misalignment to the horizon line makes a box look subtly "off" or deformed, but beginners don't know why). Generic LLMs hallucinate visual measurements: If you ask a multi-modal LLM to critique a drawing, it often invents arbitrary numbers ("your angle is off by 15 degrees") without any spatial ground truth. To bridge this gap, I designed Atelier around a single strict invariant: "The geometry measures, the AI teaches, the student grows." (ADR-001) Atelier is an agentic AI studio master for remote art students that decouples deterministic geometric calculation (via OpenCV) from high-empathy pedagogical critique (via Gemini Flash on Vertex AI). ATELIER ARCHITECTURE ┌──────────────────────┐ ┌────────────────────────┐ ┌───────────────────────┐ │ Atelier.Web │ │ Gemma Router │ │ OpenCV Geometry │ │ (Blazor / .NET 10) │ ──────> │ (Vertex AI 2B/9B) │ ──────> │ - Deskew / Canny │ │ - Multimodal UX │ │ - Exercise Classify │ │ - RANSAC VPs (k=1,2) │ │ - Annotated Overlay│ │ - Pre-Gemini Tuning │ │ - Horizon Line (LH) │ └──────────────────────┘ └────────────────────────┘ │ - Error in Degrees │ │ └───────────────────────┘ │ │ ▼ ▼ ┌──────────────────────┐ ┌────────────────────────┐ ┌───────────────────────┐ │ Collaborative Loop │ <────── │ Anti-Hallucination │ <────── │ Gemini Flash Studio │ │ - ASK Clarification │ │ Validator (ADR-001) │ │ (Google Vertex AI) │ │ - GUIDE Next Drill │ │ - Verifies degrees │ │ - Level-Aware Rubric │ │ - CAPTURE Feedback │ │ - Rejects fakes │ │ - Two-Plane Critique │ │ - ADAPT Profile │ └────────────────────────┘ └───────────────────────┘ Atelier divides every critique into two distinct, validated planes: Plane A (Measured Findings): 100% strictly derived from OpenCV metrics. It reports exact vanishing points ($VP$ or $F_1, F_2$), horizon line tilt, and per-line angular deviation in degrees. Plane B (Studio Observations): Qualitative criteria where the LLM evaluates the drawing like a human master instructor (line weight contrast between construction traces and definitive contours, spatial legibility, and cleanliness). An Anti-Hallucination Validator intercepts the critique. If the model mentions any numerical measurement not found in the OpenCV payload, the response is rejected and regenerated with corrective feedback. Atelier is not a one-shot chatbot; it acts as a proactive studio partner orchestrating 4 collaborative verbs: ASK: Before analyzing, Atelier asks the student: "What were you practicing today? Which part felt hardest?" The answers calibrate the feedback depth. GUIDE: Instead of generic advice, Atelier prescribes specific follow-up drills targeted at the student's primary recurring deviation (e.g., Targeted $F_1$ Convergence Drill). CAPTURE: After every review, the student provides explicit feedback (helpful: bool + note). This is persisted as an immutable event in Google Cloud Firestore (ADR-005). ADAPT: Learning profiles are never edited manually. Atelier dynamically derives the student's tone_preference (adapting from technical to encouraging if recent feedback indicates frustration) and tracks the convergence error reduction curve over time. Async-first Ingestion: Students or parents drop sketchbook photos into a private Google Cloud Storage bucket (atelier-inbox/{studentId}/). Eventarc fires a CloudEvent to the Cloud Run agent service, executing geometry calculation and critique in the background. Weekly Digest: Cloud Scheduler triggers weekly aggregations, calculating the error reduction percentage and prescribing a 3-day practice plan (Monday, Wednesday, Friday) for the upcoming week. Production Hardening: Deployed on Google Cloud Run with .NET 10 (using KnownIPNetworks.Clear() for proxy header termination) and Python FastAPI microservices. Building Atelier demonstrated that the future of agentic AI in technical disciplines (art, engineering, surgery, architecture) requires hybrid intelligence: Use computer vision for math, physics, and ground truth. Use generative LLMs for language, empathy, pedagogy, and inspiration. When these two forces combine, students of all ages can see their invisible mistakes and grow with confidence. Built with Google Agent Development Kit (ADK), Vertex AI (Gemini Flash & Gemma), Google Cloud Run, Cloud Storage, Eventarc, Firestore, OpenCV, and .NET 10.