Concierge AI is a next-generation intelligent routing platform that revolutionizes how users get tax and financial help. By combining advanced AI with human expertise, it instantly analyzes every question and routes it to the perfect solution: lightning-fast AI responses for common queries or personalized expert assistance for complex scenarios.
Built with cutting-edge semantic routing, real-time RAG (Retrieval Augmented Generation), and intelligent expert matching, Concierge AI delivers the right answer, from the right source, at exactly the right time.
- Semantic Intent Detection: Advanced NLP models automatically classify user questions into categories like Simple Tax, Complex Tax, or Urgent scenarios
- Dynamic Complexity Scoring: Real-time analysis determines if a query needs AI assistance or human expertise
- Context-Aware Routing: Considers query complexity, confidence levels, and user intent for optimal routing decisions
- AI Agent Mode: Handles routine tax questions instantly using RAG-powered responses grounded in a comprehensive vector knowledge base
- Expert Matching System: Automatically matches complex queries to the best-suited human expert based on specialization, availability, and match confidence
- Seamless Handoff: Smooth transitions between AI and human assistance ensure users always get the best possible help
- Vector Knowledge Base: Tax documents and financial information stored in Supabase with pgvector for ultra-fast semantic search
- Contextual Retrieval: Each AI response pulls relevant source documents in real-time to provide accurate, grounded answers
- Source Attribution: Every AI answer includes references to source documents for transparency and trust
- Glassmorphic UI: Modern, responsive interface with smooth animations and premium design aesthetics
- Real-Time Feedback: Live analysis indicators show users how their query is being processed
- Mobile-First Design: Fully responsive experience optimized for all device sizes
- Expert Profiles: Rich expert cards with bios, specializations, and availability status
- Next.js 14: Latest App Router with server components for optimal performance
- TypeScript: Type-safe development with full IDE support
- Tailwind CSS: Utility-first styling with custom design system
- Framer Motion: Smooth, professional animations and transitions
- Lucide Icons: Modern, consistent icon library
- FastAPI: High-performance Python API framework with automatic documentation
- LiteLLM: Unified LLM interface (serverless optimized)
- Semantic Router: Intent classification and query routing engine
- Sentence Transformers: State-of-the-art embeddings for semantic search
- Groq LLM: Ultra-fast inference with Llama 3 models
- Supabase: Serverless PostgreSQL with pgvector for vector operations
- Vercel: Edge-optimized hosting with automatic deployments
- Python Serverless Functions: Backend deployed as Vercel serverless functions
- Environment Management: Secure secrets handling across environments
concierge-ai/
├── app/ # Next.js App Router pages
│ ├── page.tsx # Landing page with hero, features, architecture
│ └── chat/ # Chat interface with AI/expert routing
├── components/
│ ├── main/ # Main page components (Hero, Features, Footer)
│ ├── ui/ # Reusable UI components (Buttons, Cards, Effects)
│ └── sub/ # Sub-components and utilities
├── api/ # FastAPI backend
│ ├── services/ # Core business logic (RAG, routing, matching)
│ ├── utils/ # Helper utilities (embeddings, database)
│ └── index.py # Main API entry point
├── lib/ # Frontend utilities and API clients
└── knowledge_data/ # Tax knowledge documents for RAG
- Node.js 18+ - JavaScript runtime
- Python 3.9+ - Backend runtime
- Supabase Account - For database and vector storage
- Groq API Key - For LLM inference
- Hugging Face Token - For embeddings API
-
Clone the repository
git clone https://github.com/rshriharripriya/concierge-ai.git cd concierge-ai -
Install frontend dependencies
npm install
-
Set up Python environment
python -m venv .venv source .venv/bin/activate # On Windows: .venv\Scripts\activate # For Production/Vercel (Lightweight): pip install -r requirements.txt # For Local Dev & Eval (Includes Ragas/LangChain): pip install -r requirements-dev.txt
-
Configure environment variables
Create
.env.localin the project root:# Supabase Configuration SUPABASE_URL=your_supabase_project_url SUPABASE_KEY=your_supabase_service_role_key NEXT_PUBLIC_SUPABASE_URL=your_supabase_project_url NEXT_PUBLIC_SUPABASE_ANON_KEY=your_supabase_anon_key # AI Service Keys GROQ_API_KEY=your_groq_api_key HF_TOKEN=your_huggingface_token # API Configuration NEXT_PUBLIC_API_URL=/api/py
Where to get these keys:
- Supabase: Sign up at supabase.com → Create project → Settings → API
- Groq: Get free API key at console.groq.com
- Hugging Face: Sign up at huggingface.co → Settings → Access Tokens
-
Initialize the database
Run the Supabase schema setup:
# Execute the SQL in your Supabase dashboard cat supabase_schema.sql -
Start development servers
# Frontend (runs on http://localhost:3000) npm run dev # Backend (in a separate terminal) #Remove conda environment ( if using venv) conda deactivate source .venv/bin/activate cd concierge-ai python -m uvicorn api.index:app --reload --port 8000
The semantic router analyzes natural language queries to determine the best handling path. It evaluates:
- Intent classification (Simple Tax, Complex Tax, Investment, etc.)
- Complexity scoring (1-5 scale based on query difficulty)
- Confidence levels (How certain the system is about classification)
- Routing decisions (AI vs Human expert)
When a query routes to the AI agent:
- Query is converted to embeddings using Sentence Transformers
- Vector similarity search finds the most relevant tax documents
- Retrieved context is sent to Groq Llama 3 for response generation
- AI synthesizes an answer grounded in actual tax documentation
- Response includes source citations for transparency
For complex queries that need human expertise:
- Query embeddings are compared against expert skill profiles
- Semantic similarity scores determine best expert matches
- Availability and match confidence are factored in
- User sees expert profile with bio and specialization
- Seamless handoff to human expert occurs
- Glassmorphism Design: Modern frosted glass effects with blur and transparency
- Animated Beams: Dynamic system architecture visualization with flowing data
- Smooth Transitions: Framer Motion animations for premium feel
- Responsive Layout: Mobile-first design that adapts to any screen size
- Dark Mode Ready: Elegant crimson theme with professional aesthetics
- Server Components: Next.js server components reduce client bundle size
- Code Splitting: Automatic route-based code splitting
- Vector Search: Sub-100ms semantic search with pgvector
- Edge Deployment: Hosted on Vercel edge network for global low latency
- Lazy Loading: Components and images load on-demand
- Environment Variables: Sensitive keys never exposed to client
- Type Safety: Full TypeScript coverage prevents runtime errors
- SQL Injection Protection: Parameterized queries via Supabase client
- CORS Configuration: Proper cross-origin resource sharing
- API Rate Limiting: Protection against abuse
The system includes pre-loaded tax knowledge covering:
- Standard deductions and tax brackets
- International student tax obligations
- Investment and capital gains taxation
- Tax credits and deductions
- Filing extensions and deadlines
This is a portfolio project showcasing advanced AI engineering skills. Feel free to fork and build upon it!
MIT License - See LICENSE file for details