--- title: loopbreakers emoji: 🧠 colorFrom: purple colorTo: purple sdk: docker pinned: true tags: - flask - claude - math - education - tutoring - ai - deepsite - frontend - anthropic --- # LoopBreakers – Reward-Hacking Resistant AI Math Tutor LoopBreakers is an AI-powered web app designed to provide **clear, step-by-step math solutions** for Year 5–10 students. It features **reward-hacking resistance logic**, powered by **Anthropic's Claude-3**, and demonstrates how feedback-driven alignment strategies can improve answer quality in real-time. ## 🔍 What it does - Solves K–12 math word problems via Claude-3 API - Extracts and verifies final numeric answers - Automatically retries with improved prompting if initial output is unclear or wrong - Displays detailed, alignment-aware explanations for educational transparency ## 🛠️ Technologies - Claude-3 Opus API (`claude-3-opus-20240229`) - Python + Gradio frontend - Hosted on Hugging Face Spaces - UI generated with [DeepSite](https://enzostvs-deepsite.hf.space/) ## ✅ Why it matters Unlike traditional LLM math apps, **LoopBreakers is built to avoid shortcut exploitation**. The retry logic allows users to see **how aligned prompting improves reasoning**, highlighting real-time effects of AI reward optimization in education. ## 📥 Usage 1. Type any math word problem into the input box. 2. Click **"Get Solution"**. 3. View the **step-by-step explanation** and the **final answer**. 4. If the answer was incorrect, the AI will **automatically retry with better alignment logic.** > Example: > _"Maya has 5 boxes. Each box contains 3 red balls and 2 blue balls. She gives 2 boxes to her friend. How many red balls does Maya have left?"_ --- **Made with 💡 using Claude + DeepSite by the LoopBreakers capstone team.** Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference