PRODUCTION CASE STUDY // BUILT AT BHANZU Feb 2026 - Apr 2026

AI Math Tutor: Guided Doubt Solving

Orchestrating Gemini 3.1 Pro & Flash for real-time visual mistake highlighting and step-by-step feedback

Pedagogical Scaffolding & Bounding Box Sync

To solve the tendency of standard Gen-AI wrappers to simply spoil math answers for students, we engineered a spatial AI Math Tutor. By routing cognitive logic selectively between Gemini 3.1 Pro and Flash, the system acts like a human teacher—delivering real-time step-by-step feedback through chat/talk and circling errors on their work canvas rather than handing out final solutions. Gemini 3.1 was selected as the foundation for this system due to its excellent handwriting recognition skills, allowing the models to parse highly diverse, non-standard student ink strokes and mathematical writing with absolute precision.

1. Student Input & Parallel Dispatch

As the student works on the Canvas UI (writing equations or asking questions), their stroke data and chat messages stream directly to the centralized Tutor Orchestrator, which dispatches parallel queries to Gemini 3.1 Flash and Gemini 3.1 Pro at the same time.

2. Immediate Feedback Stream (Gemini 3.1 Flash)

The Gemini 3.1 Flash model returns quickly (in 3-4 seconds) with a streaming type response, providing near-instant conversational feedback, receipt confirmation, or quick cues so the student never feels stuck or waiting.

3. Deep Logical Evaluation (Gemini 3.1 Pro)

In parallel, the high-reasoning Gemini 3.1 Pro model processes the logical and structural algebraic progression in the background, taking a little longer (10-12 seconds) to run a comprehensive validation check.

4. Step-by-Step Feedback & Red Circle Highlight

Soon after, Gemini Pro's logical assessment returns. If a step error is spotted, the orchestrator delivers targeted step-by-step feedback to the UI and triggers the Canvas to draw a red circle directly around the incorrect term to guide the student toward self-correction.

Visual Request & Evaluation Pipeline

Interactive live view of data flows, vision OCR coordinates, and logic checks navigating the tutor system.

🤖

Animated request flow visualization loading...

Key Interaction Features

Creating a natural, non-intrusive learning experience with real-time step-by-step feedback required designing specialized canvas rendering and multimodal vision components:

📷

Physical Worksheet OCR Scanner

Instead of typing complex mathematical notation, students scan physical worksheets. Leveraged by Gemini 3.1's excellent handwriting recognition skills, the system processes the raw vision stream to align coordinate grids and accurately map hand-written formulas directly to structured LaTeX nodes on the digital canvas.

🔴

Interactive Error Circling

A custom visual renderer maps bounding coordinates returned by the Gemini vision evaluation model directly to screen pixels. It overlays a hand-drawn-style red circle animation around the incorrect operator or number in real-time.

Rigorous Unit Economics

To scale 1:1 tutoring globally, the orchestrator routes simple transcription and UI tasks to the rapid Gemini 3.1 Flash model, reserving the more expensive Gemini 3.1 Pro logic evaluator solely for logic checks.

Operation Cost Vector Gemini 3.1 Flash Routing Gemini 3.1 Pro Routing
Vision OCR / Digitization ₹0.04 per page scan
Processes image coordinates
N/A (Routed to Flash)
Chat & Dialogue Scaffolding ₹0.02 per query
Handles conversational phrasing
N/A (Routed to Flash)
Logical Step Evaluation N/A (Requires high-reasoning) ₹0.38 per logic check
Verifies mathematical logic integrity
Blended Cost per Step ~₹0.12 average cost per problem-solving step (vs ₹300/hr human coach)

Business Impact Summary

92% Student Self-Correction Rate

Students successfully resolve flagged errors on their own without needing the tutor to reveal the correct answer.

65% Support Ticket Scaling Cost Reduction

Direct savings in student dashboard support operations by shifting 1:1 live doubt-clearing calls to automated scaffolding.