## I built this project to showcase my passion for building end-to-end AI systems. **Coach Aria** isn't just another chatbot — it's a fully-functional, multimodal AI agent that can: - **Have meaningful coaching conversations** powered by Llama 3.3 - **Understand your voice** using Whisper STT - **Understand images you send** using Llama 4 Scout Vision - **Speak back to you** with natural voice synthesis via ElevenLabs - **Generate contextual images** using Stable Diffusion XL - **Remember important details** about you across sessions using vector memory - **Intelligently route** between text, audio, and image responses > I built this to demonstrate that I can design, architect, and implement complex AI systems from scratch — not just follow tutorials. ## Features in Action ### Voice Understanding Aria can listen to your voice messages and transcribe them in real-time using Whisper.

*Voice messages are transcribed using Whisper Large v3 Turbo via Groq* ### Image Understanding Aria can analyze and understand images you send, providing detailed descriptions and context-aware responses.

*Aria uses Llama 4 Scout Vision model to understand and describe images* --- ### Voice Response Aria can speak back to you with natural, expressive voice synthesis. https://github.com/BhimPrasadAdhikari/life-coach/raw/main/public/screenshots/videos/agent_speak.mp4 *Natural voice responses powered by ElevenLabs TTS* --- ### AI Image Generation Aria can generate contextual images based on the conversation. Here are some examples:

*Images generated by Stable Diffusion XL based on coaching conversations* --- ## What Can Aria Do? | Capability | Description | Technologies | |------------|-------------|--------------| | **Conversational Coaching** | Empathetic, growth-mindset based conversations | LangGraph, Groq Llama 3.3 | | **Voice Input** | Transcribes your voice messages in real-time | Whisper Large v3 Turbo via Groq | | **Voice Output** | Sends voice responses that sound natural | ElevenLabs TTS | | **Image Understanding** | Analyzes images you send and responds intelligently | Llama 4 Scout Vision via Groq | | **Image Generation** | Creates visualizations based on conversation context | Stable Diffusion XL via HuggingFace | | **Long-Term Memory** | Remembers your goals, challenges, and preferences | Vector Store + Semantic Search | | **Smart Routing** | Automatically decides when to respond with text, voice, or images | LLM-powered intent classification | --- ## System Architecture ### High-Level Workflow ```mermaid graph TD A[User Input] --> B[Memory Retrieval] B --> C[Smart Router] C -->|Text| D[Conversation Node] C -->|Audio| E[Audio Node - TTS] C -->|Image| F[Image Node - SDXL] D --> G[Memory Saving] E --> G F --> G G --> H[Response to User] ``` ### Detailed Node Operations ```mermaid flowchart TB subgraph INPUT["Input Processing"] UI[User Message] --> |Text/Voice/Image| PRE[Preprocessing] PRE --> |Voice| STT[Speech-to-Text
Whisper v3] PRE --> |Image| ITT[Image-to-Text
Llama 4 Scout] STT --> MSG[Processed Message] ITT --> MSG PRE --> |Text| MSG end subgraph MEMORY["Memory Layer"] MSG --> MR[Memory Retrieval Node] MR --> |Semantic Search| VS[(Qdrant
Vector Store)] VS --> |Relevant Memories| CTX[Context Injection] end subgraph ROUTING["Smart Routing"] CTX --> ROUTER[Router Node
LLM Classification] ROUTER --> |conversation| CONV ROUTER --> |audio| AUDIO ROUTER --> |image| IMG end subgraph RESPONSE["Response Generation"] CONV[Conversation Node
Llama 3.3 70B] AUDIO[Audio Node
+ ElevenLabs TTS] IMG[Image Node
+ SDXL Generation] end subgraph OUTPUT["Output & Learning"] CONV --> MS[Memory Saving Node] AUDIO --> MS IMG --> MS MS --> |Extract Important Info| VS MS --> RESP[Response to User] end ``` ### How Each Node Works | Node | Purpose | What It Does | |------|---------|--------------| | **Memory Retrieval** | Context Loading | Searches vector store for relevant past memories using semantic similarity | | **Router** | Intent Classification | Uses LLM to classify if user wants text, audio, or image response | | **Conversation** | Text Response | Generates empathetic coaching response using Llama 3.3 | | **Audio** | Voice Response | Generates text response + converts to speech via ElevenLabs | | **Image** | Visual Response | Creates scenario prompt → enhances it → generates image via SDXL | | **Memory Saving** | Learning | Analyzes conversation for important facts and stores in vector DB | --- ## Tech Stack | Layer | Technology | |-------|------------| | **Framework** | LangGraph | | **LLM** | Groq (Llama 3.3 70B) | | **Vision** | Llama 4 Scout Vision | | **STT** | Whisper Large v3 Turbo | | **TTS** | ElevenLabs | | **Image Gen** | Stable Diffusion XL | | **Vector DB** | Qdrant | | **Interface** | Chainlit | | **Deployment** | Google Cloud Run | --- ## Getting Started ### Prerequisites - Python 3.13+ - API Keys for: Groq, ElevenLabs, HuggingFace, Qdrant ### Installation ```bash # Clone the repository git clone https://github.com/BhimPrasadAdhikari/whatsapp-agent.git cd whatsapp-agent # Create virtual environment python -m venv venv source venv/bin/activate # On Windows: venv\Scripts\activate # Install dependencies pip install -r requirements.txt ``` ### Environment Setup Copy the example environment file and add your API keys: ```bash cp .env.example .env ``` Then edit `.env` with your API keys. **📖 See [API Setup Guide](docs/API_SETUP_GUIDE.md) for detailed instructions on obtaining all API keys.** ```env GROQ_API_KEY=your_groq_api_key ELEVENLABS_API_KEY=your_elevenlabs_api_key HF_TOKEN=your_huggingface_token QDRANT_URL=your_qdrant_url QDRANT_API_KEY=your_qdrant_api_key ``` ### Run the Application ```bash chainlit run interfaces/chainlit/app.py -w ``` --- ## Let's Connect! I'm actively looking for **internship and full-time opportunities** in AI/ML Engineering.

LinkedIn •

--- ## License This project is open source and available under the [MIT License](LICENSE). ---

Built with ❤️ by Bhim Prasad Adhikari

AI&ML based projects I've built

Multi-model AI life coach agent.