EvolvApp

Inspiration

In today’s AI landscape, most tools generate one output per prompt. But real creativity doesn’t work like that — humans explore, compare, refine, and evolve ideas.

We were inspired by a simple question: “What if AI didn’t just generate answers… but explored entire spaces of possibilities?”

By combining evolutionary algorithms (MAP-Elites) with modern AI models like GLM, we wanted to move from generation → exploration → evolution.

What it does

EvolvApp is an AI-powered system that turns a single idea into a diverse, evolving ecosystem of working applications.

Instead of generating one app, it:

Creates a 3×3 grid of apps across different styles and complexities Each app is fully functional (HTML/CSS/JS) Lets users preview, compare, and select favorites

Then it goes further:

⚡ Mutation → evolve apps from a selected favorite 🧬 Crossover → combine design from one app and features from another 🔁 Continuous evolution → explore better versions over generations

It transforms app creation into an interactive evolutionary process.

How we built it

We built EvolvApp as a full-stack AI agent system with a frontend-first architecture:

Frontend: React + Vite for fast, interactive UI AI Integration: GLM model via API (OpenRouter / Z.ai) Core Engine: MAP-Elites algorithm for diversity-based search Grid-based archive (3×3 niches) Agents inside the system: Generator Agent → builds initial apps Mutation Agent → evolves variations Crossover Agent → merges designs + features Selection Agent → guided by user preference

Each app is generated as a self-contained HTML file and rendered in real-time using iframes.

Challenges we ran into

API reliability & rate limits → Solved with batching, retries, and multi-provider support Balancing speed vs quality → Faster prompts reduced quality, so we optimized parallel generation instead Maintaining valid HTML output → Built robust extraction and fallback mechanisms Designing meaningful diversity → Carefully engineered “behavior dimensions” (style × complexity) Making evolution intuitive for users → Simplified complex concepts like mutation and crossover into clickable actions

Accomplishments that we're proud of

Built a fully working evolutionary AI system in limited time Achieved real diversity across generated apps, not just minor variations Implemented mutation + crossover, rarely seen in AI app builders Created a visually compelling demo (grid filling in real-time) Successfully turned a research concept (MAP-Elites) into a usable product

What we learned

AI is powerful, but true value comes from exploration, not single outputs Evolutionary algorithms pair extremely well with LLMs UX is critical — even advanced systems must feel simple Parallel generation strategies can significantly improve performance Combining human preference + AI evolution leads to better outcomes

What's next for EvolvApp

📄 Paper → App Mode (generate tools from research papers) ⭐ User feedback learning (fitness scoring & ranking) 🧬 Multi-parent crossover (combine multiple apps) 💾 Version history & evolution tracking ☁️ One-click deployment (GitHub/Vercel integration) 🤖 Autonomous agents that iterate without user input

Built With

• glm
• javascript
• map-elites
• react