Inspiration
What if a single prompt could generate a production-ready, exportable full-stack app — instantly?
What it does
🧠 Plans a structured application schema (entities, fields, relationships) ⚙️ Generates an Express + SQLite backend 🎨 Generates a Next.js frontend with dynamic CRUD UI 🔄 Automatically connects frontend and backend 🧪 Adds validation and safe defaults 📦 Exports a ready-to-run project 🛡 Falls back intelligently between Gemini → OpenAI → local templates
How we built it
StackPilot AI is built as a multi-agent system:
🧠 Planner Agent Uses Gemini as primary LLM Falls back to OpenAI if Gemini fails Falls back to deterministic local templates if AI is unavailable Outputs strict JSON AppPlan schema
🏗 Backend Agent Generates Express + SQLite backend Auto-creates dynamic CRUD routes per entity Safe SQL parameterization No-cache dev configuration for hackathon demos
🎨 Frontend Agent Generates Next.js 15 application Dynamic entity dashboard Full CRUD UI with drawers and modals IDE-style project preview Clean, production-like UI
📦 Export System Zips the generated workspace Provides downloadable full project Includes .env.example and clear README
🔄 Intelligent AI Orchestration Gemini primary OpenAI backup Local deterministic fallback Real-time UI indicators show which provider was used
Challenges we ran into
LLMs sometimes returned partial JSON or invalid structures. We solved this with: Strict schema validation Code fence stripping Safe JSON parsing recovery Provider fallback logic
Multi-Provider Fallback Ensuring seamless switching between Gemini and OpenAI required: Dynamic imports to prevent build failures Clean error bubbling Explicit provider metadata tracking
Generated Project Stability We encountered: npm recursive install issues, Windows CRLF conflicts, Node module resolution problems, Zip streaming errors. Each was resolved by improving dependency handling and export architecture. Enough production realism to demonstrate feasibility
Accomplishments that we're proud of
✅ Fully working multi-agent architecture ✅ Real fallback chain: Gemini → OpenAI → Local ✅ Dynamic full-stack CRUD generation ✅ Clean, professional UI (not a toy prototype) ✅ Exportable production-ready codebase ✅ Robust logging and agent status tracking ✅ Works even if all AI providers fail
What we learned
AI generation needs strong guardrails. Fallback logic is just as important as intelligence. Deterministic templates are critical safety nets. Clean UX matters — judges evaluate polish as much as technical depth. Multi-agent systems require careful orchestration and state management.
What's next for StackPilot AI
AI generation needs strong guardrails. Fallback logic is just as important as intelligence. Deterministic templates are critical safety nets. Clean UX matters — judges evaluate polish as much as technical depth. Multi-agent systems require careful orchestration and state management.
Log in or sign up for Devpost to join the conversation.