Serenova - AI Mental Health Companion

Serenova is an AI-powered mental health companion that provides supportive conversations and emotional guidance anytime.

Comment

Serenova — AI Mental Health Companion

About the Project

Serenova is a voice-first AI mental health companion powered by Amazon Nova models on AWS Bedrock. It provides empathetic, real-time support through natural voice conversation, text chat, emotion detection, and crisis safety features.

Inspiration

Mental health support is hard to access. Long wait times, high costs, and stigma keep millions from getting help. We wanted to build something that's always available — a warm, non-judgmental companion you can talk to anytime.

Amazon Nova Sonic's speech-to-speech capability inspired us to go voice-first. Voice is the most natural way humans express emotion. A text chatbot can miss the trembling in someone's voice or the long pauses between words. Nova Sonic processes audio directly, preserving those emotional signals that matter most in mental health conversations.

The name Serenova combines "Serene" (calmness, peace of mind) and "Nova" (new beginning) — representing a new path to inner peace through AI-guided support.

What It Does

Serenova offers four support modes tailored to different needs:

  • Companion — warm, empathetic conversation for everyday support
  • Crisis Support — calm, safety-focused responses with automatic crisis resource display (988 Lifeline, Crisis Text Line)
  • CBT Therapy — cognitive behavioral techniques to identify and reframe thought patterns
  • Emotion Regulation — guided breathing exercises (4-4-6 pattern) and grounding techniques

Key features:

  • Voice orb interface with real-time mic level visualization
  • Browser-native speech recognition for voice-to-text
  • AI responses powered by Amazon Nova Pro via Bedrock Converse API
  • Three-tier graceful degradation (Voice → Text → Demo mode)
  • Crisis keyword detection with automatic resource modal
  • Premium dark theme UI with glassmorphism and ambient animations
  • Four voice profiles: Empathetic, Calm, Professional, Warm

How We Built It

Frontend

React app with a custom voice orb UI. The orb uses CSS animations to reflect its state — idle, listening, thinking, responding. We used the Web Audio API's AnalyserNode for real-time mic level monitoring and the browser's SpeechRecognition API for speech-to-text transcription.

Backend

FastAPI (Python 3.11) with two versions:

  • Local dev (src/api/main.py) — full backend with Nova Sonic voice streaming, WebSocket support, MCP tools, emotion detection, and crisis detection
  • Production (backend/main.py) — standalone, minimal backend for App Runner deployment using Nova Pro text chat via the Bedrock Converse API

AI Models

  • Amazon Nova Sonic (amazon.nova-sonic-v1:0) — speech-to-speech voice conversation
  • Amazon Nova Pro (amazon.nova-pro-v1:0) — text chat, reasoning, response generation
  • Amazon Nova Lite (amazon.nova-lite-v1:0) — fast lightweight text responses

Deployment

  • Backend: Docker image → Amazon ECR → AWS App Runner
  • Frontend: AWS Amplify (auto-deploy from GitHub)
  • Infrastructure: AWS CDK for DynamoDB, S3, EventBridge, SNS, CloudWatch

Key Technical Decisions

  1. Voice-first with text fallback — Nova Sonic for the full experience, Nova Pro as a reliable fallback when voice isn't available
  2. Docker for App Runner — source-based Python deployments had PATH issues with uvicorn; Docker solved it completely
  3. Browser SpeechRecognition — instead of building a custom STT pipeline, we leveraged the browser's built-in speech recognition for the text fallback mode
  4. Standalone backend for production — separated backend/main.py from the full src/api/main.py to avoid complex import chains and Nova Sonic SDK dependencies on App Runner
  5. In-memory session history — kept it simple with a 20-message sliding window per session, no database needed for the MVP
  6. Glassmorphism dark theme — mental health apps need to feel calm and safe; the dark theme with soft glows and ambient animations creates that atmosphere

Challenges We Ran Into

App Runner Deployment (The Big One)

App Runner's source-based Python runtime installs packages during build but runs in a separate environment where pip-installed binaries like uvicorn aren't in PATH. We tried:

  • uvicorn main:app — "executable not found"

The fix: switched to Docker image deployment via ECR. Building our own Docker image meant uvicorn was properly installed and in PATH. Problem solved.

Nova Sonic SDK Dependencies

The Nova Sonic SDK has heavy dependencies that aren't needed for text-only deployment. We solved this with lazy imports (try/except ImportError) and a separate standalone backend/main.py that only uses boto3 for the Bedrock Converse API.

Frontend Build on Amplify

npm ci failed because package-lock.json had a TypeScript version mismatch. Fixed by regenerating the lock file with npm install.

CORS and API Integration

Getting the React frontend to talk to the FastAPI backend across different origins required careful CORS configuration. We used allow_origins=["*"] for development simplicity.

Accomplishments That We're Proud Of

  • Working live deployment — backend on App Runner, frontend on Amplify, both running in production
  • Three-tier degradation — the app never crashes; it gracefully falls back from voice to text to demo mode
  • Voice orb UI — the animated orb with real-time mic levels feels premium and calming
  • Crisis detection — automatic keyword detection with immediate display of 988 Lifeline and other resources
  • Breathing exercise — the 4-4-6 guided breathing with animated circle actually helps calm you down
  • Four support modes — each mode has a distinct AI personality tuned for its purpose
  • Full documentation — architecture diagrams, deployment guide, user guide, and this writeup

What We Learned

  • App Runner source-based Python deployments are unreliable — Docker/ECR is the way to go
  • Nova Sonic is powerful but requires careful integration — the streaming API needs proper session management and error handling
  • Browser APIs are underrated — SpeechRecognition and Web Audio API gave us voice features without any backend dependencies
  • Graceful degradation is essential — users should never see a broken app, just a reduced-feature version
  • Mental health AI needs careful prompt engineering — the system prompts for each mode required iteration to feel warm without being patronizing
  • Keep production deployments simple — the standalone backend/main.py with minimal dependencies was the key to a successful deployment

What's Next for Serenova

  • Full Nova Sonic voice streaming in production — deploy the complete voice pipeline on App Runner or ECS
  • Persistent conversation history — DynamoDB storage for cross-session memory
  • Emotion detection from audio — use Nova Sonic's audio features (pitch, volume, pace) for real-time emotion analysis
  • Mobile app — React Native version for iOS and Android
  • Multi-language support — Nova Sonic supports multiple languages
  • Therapist dashboard — optional professional oversight with user consent
  • Journal feature — AI-guided journaling with mood tracking over time
  • Group support sessions — moderated AI-assisted group conversations

Built With

Category Technologies
AI Models Amazon Nova Sonic, Amazon Nova Pro, Amazon Nova Lite (AWS Bedrock)
Backend Python 3.11, FastAPI, Uvicorn, boto3
Frontend React 18, JavaScript, CSS3, Web Audio API, SpeechRecognition API
Deployment AWS App Runner, AWS Amplify, Amazon ECR, Docker
Infrastructure AWS CDK, DynamoDB, S3, EventBridge, SNS, CloudWatch, IAM
Design Glassmorphism, CSS animations, dark theme
Tools Git, GitHub, AWS CLI, Docker Desktop

Serenova is an AI companion and is not a replacement for professional mental health care. If you or someone you know is in crisis, call 988 or text HOME to 741741.

Built With

Share this project:

Updates

posted an update —

Serenova – AI Mental Health Companion

We’re excited to share progress on Serenova, an AI-powered companion designed to provide supportive conversations and emotional guidance.

The application uses Amazon Bedrock with Nova AI models to generate empathetic responses and is built with React, FastAPI, and AWS cloud services including AWS App Runner, Amazon ECR, and AWS Amplify.

Serenova supports voice and text interactions, helping users express their thoughts and practice calming techniques in a safe environment.

Built by Dineshraj Dhanapathy, Sakshi Semalti, and Swetha Mudunuri with the goal of using AI and cloud technology to support mental wellbeing.

Log in or sign up for Devpost to join the conversation.