AI App Compiler

Overview

AI App Compiler converts natural language software requirements into structured application configurations using Gemini-powered intent extraction and a multi-stage compiler architecture.

Example Input:

Build a CRM with login, contacts, dashboard, role-based access, and premium plan.

Example Output:

• UI Schema
• API Schema
• Database Schema
• Authentication Rules
• Validation Results
• Repaired Configuration

---

Architecture

User Prompt

↓

Intent Extraction (Gemini 2.5 Flash)

↓

System Design

↓

Schema Generation

↓

Validation

↓

Repair

↓

Runtime Simulation

↓

Final Structured Output

---

Pipeline Stages

1. Intent Extraction

Converts natural language requirements into a structured intermediate representation using Gemini 2.5 Flash.

2. System Design

Generates entities, pages, user roles, and application architecture.

3. Schema Generation

Generates:

• UI Schema
• API Schema
• Database Schema
• Authentication Rules

4. Validation

Checks:

• Missing keys
• Invalid structures
• Schema consistency
• API ↔ Database consistency
• Role validation

5. Repair Engine

Automatically repairs known schema issues without regenerating the entire output.

6. Runtime Simulation

Simulates execution readiness and validates generated configurations.

---

Failure Handling

The system handles:

• Vague prompts
• Underspecified requirements
• Invalid configurations
• Conflicting requirements

Example 1: Vague Prompt

Input:

Build app

Output:

{
  "clarification_needed": true,
  "message": "Prompt is too vague. Please provide more details."
}

Example 2: Conflicting Requirements

Input:

Build an ecommerce app without products

Output:

{
  "conflict_detected": true,
  "message": "Conflicting requirements: ecommerce applications require products."
}

---

Example AI Output

Input:

Build a gym management SaaS with trainers, members, subscriptions and attendance tracking.

Output:

{
  "app_type": "Gym Management SaaS",
  "features": [
    "Trainer Management",
    "Member Management",
    "Subscription Management",
    "Attendance Tracking"
  ],
  "roles": [
    "Admin",
    "Trainer",
    "Member"
  ]
}

---

Evaluation Framework

Dataset:

• 10 real product prompts
• 10 edge-case prompts

Metrics:

• Success Rate
• Validation Errors
• Repair Attempts
• Failure Types

---

Tech Stack

Backend

• FastAPI
• Python

AI

• Gemini 2.5 Flash

Frontend

• Next.js
• React
• Tailwind CSS

---

Project Structure

AI-App-Compiler/
│
├── backend/
│   ├── stages/
│   ├── validators/
│   ├── repair/
│   └── runtime/
│
├── frontend/
│
├── evaluation/
│   ├── prompts.json
│   └── metrics.md
│
└── README.md

---

Running Locally

Backend

cd backend
venv\Scripts\activate
uvicorn main:app --reload

Frontend

cd frontend
npm install
npm run dev

---

Runtime Validation

To ensure generated configurations are execution-ready, the system performs runtime simulation before accepting output.

Runtime validation verifies:

• UI pages are present
• API endpoints are generated
• Database tables exist
• Authentication roles are defined

Example Runtime Result:

{
  "executable": true,
  "pages_found": 5,
  "endpoints_found": 4,
  "tables_found": 4
}

This provides execution awareness and helps ensure generated configurations can be used by downstream application generators without manual fixes.

---

Cost vs Quality Tradeoffs

The system balances quality, reliability, latency, and inference cost.

Design decisions:

• Gemini 2.5 Flash is used for intent extraction because it provides strong reasoning quality with low latency.
• Validation and repair stages reduce the need for expensive full regeneration.
• Multi-stage generation improves reliability while keeping inference costs manageable.
• Deterministic schema generation ensures predictable outputs after AI intent extraction.

Tradeoff Summary:

Objective	Approach
Quality	Gemini-powered intent extraction
Reliability	Validation + Repair Engine
Latency	Lightweight schema generation
Cost	Single AI call followed by deterministic processing

---

Evaluation Results

Test Dataset:

• 10 real product prompts
• 10 edge-case prompts

Tracked Metrics:

• Success Rate
• Validation Errors
• Repair Attempts
• Failure Types
• Latency

Example Results:

{
  "total_tests": 20,
  "successful": 18,
  "failed": 2,
  "success_rate": "90%",
  "average_latency": "1.1s",
  "failure_types": [
    "vague_prompt",
    "conflicting_requirements"
  ]
}

---

Future Improvements

• Multi-model AI support
• Full application code generation
• Database migration generation
• Cloud deployment
• Multi-agent architecture

---

Author

Saaketh B