Mimir

Overview

Students often struggle not because they lack resources, but because learning tools don't adapt to how they think and work. We wanted to create an AI tutor that actually sees your workspace, like your notes and work, and adapts its teaching style to match your needs.

What We Built

Core Features
Multi-Modal Workspace
Text editor for notes
Code editor with syntax highlighting
Annotation canvas for sketching diagrams
Seamless switching between all three

Context-Aware AI

Type @ to mention any workspace instance in your chat AI sees your active work and referenced instances Annotation canvases exported as images for visual analysis Branching Conversations Explore multiple solution paths Tree view to visualize conversation history No lost context when exploring alternatives Manim Animation Generation Automatic visualization of math and physics concepts Request animations with keywords like "visualize" or "animate"

Tech Stack:

Frontend: Next.js 15, TypeScript, React, Excalidraw
Backend: FastAPI (Python), Anthropic Claude API
Features: Server-Sent Events for streaming, fuzzy search for mentions, Supabase for persistence Key Implementation:
Workspace context builder assembles active/mentioned instances into structured prompts
Teaching mode system injects tailored instructions into Claude's system prompt
Annotation canvas exports drawings as base64 PNGs for Claude's vision API
Intelligent content truncation manages context window limits

Challenges & Solutions

Context Window Management: Implemented smart truncation—full content for active/mentioned instances, truncated for others
Fuzzy Mention Matching: Multi-stage matching (exact → substring → character overlap) with 70% similarity threshold
Generating Manim code: Claude had difficulties generating proper Manim code, so we implemented a self testing and regeneration loop

Impact Mimir bridges the gap between traditional learning tools by:

Adapting teaching style to individual learner needs
Understanding workspace context to provide relevant help
Supporting multiple modalities (text, code, visual) in one seamless interface
Enabling exploration through branching conversations without losing context
It's not just a chatbot—it's a complete learning environment that meets students where they are.

Check out our slides and video here: https://docs.google.com/presentation/d/1nqS7gaEjR2E71aczTwY_Z7GkQX5jIpavUNu-RMiDYL0/edit?usp=sharing