Inspiration

Mental health struggles affect 1 in 5 adults, yet immediate, personalized support remains inaccessible for most people. During a particularly stressful period, I found myself searching for quick relief techniques but felt overwhelmed by generic advice that didn't fit my specific situation. I thought: "What if AI could create personalized coping strategies on-demand, like having a therapist design a custom toolkit for exactly what I'm going through?"

That's when Inner Harbor was born - an AI-powered app that generates personalized 3–5 minute "micro-rituals" for immediate emotional relief.

What it does

Inner Harbor transforms emotional distress into actionable relief through AI-generated personalized rituals:

Input
Users select their emotion (anxious, sad, overwhelmed, etc.) and context (work, relationships, health, etc.), rate their distress (0–10), and add optional notes.
AI Generation
GPT-4o-mini creates a custom ritual combining evidence-based techniques from CBT, DBT, and mindfulness—tailored to their exact situation.
Guided Experience
Users complete the ritual step-by-step with built-in timers and clear instructions.
Track Progress
Before/after distress ratings and helpfulness scores build a personalized library of what works.
Visualize Patterns
The "Emotional Weather" timeline reveals patterns and progress over time.

Critical Safety Features

Comprehensive crisis detection (37+ keywords) with immediate hotline resources
Clear disclaimers that this is NOT therapy
Privacy-first design with user data control

How we built it

Tech Stack

Frontend: Next.js 15 with TypeScript and Tailwind CSS
Backend: Next.js API Routes with Prisma ORM
Database: PostgreSQL via Supabase with connection pooling
AI: OpenAI GPT-4o-mini with custom prompt engineering
Deployment: Vercel with automatic GitHub integration and CI/CD

Development Process

Started with comprehensive product specification and task breakdown
Built core ritual flow with mock data to validate UX
Implemented database schema with Prisma (4 tables: users, sessions, rituals, runs)
Integrated AI with careful prompt engineering for safety and personalization
Added crisis detection system with extensive testing (37 scenarios)
Implemented tracking features (library, favorites, emotional weather)
Hardened for production (rate limiting, input sanitization, error handling, security headers)
Created comprehensive test suites (10/10 E2E tests, 37/37 crisis tests passing)
Deployed to production with full documentation

AI Assistant Partnership

I used Claude Code (Sonnet 4.5) as my AI pair programmer throughout development, which accelerated development while teaching me best practices for production-ready applications.

Challenges we ran into

1. IPv4/IPv6 Compatibility with Supabase

The biggest technical hurdle was database connectivity. Supabase's direct connection requires IPv6, but most networks are IPv4-only. After multiple failed attempts, I discovered I needed to use Supabase's session pooler for both app connections AND migrations, requiring special Prisma configuration.

2. Responsible AI Design

Creating an AI system for mental health required careful consideration:

How do we ensure rituals are helpful, not harmful?
How do we detect crisis situations reliably?
How do we set clear boundaries about what the app can and cannot do?

This was solved through prompt engineering, crisis keyword detection, and prominent safety disclaimers.

3. Build Errors in Production

During Vercel deployment, TypeScript compiled test scripts that shouldn’t be included. I learned to correctly configure tsconfig.json to exclude test directories.

4. Balancing Personalization with Safety

The AI must be creative yet safe. I approached this by:

Training the AI with specific therapeutic techniques (CBT, DBT, mindfulness)
Enforcing strict time limits and step counts
Validating ritual structure before showing users
Implementing retry logic and fallback rituals

Accomplishments that we're proud of

Fully Production-Ready: Live at https://inner-harbor-eta.vercel.app
100% Test Pass Rate: 10/10 E2E tests and 37/37 crisis detection tests
Complete Full-Stack Application
Responsible AI with crisis detection and safety systems
Real-World Impact with personalized, effective rituals
Production Quality:
- Rate limiting (10 rituals/min/IP)
- Input sanitization and validation
- Security headers (CSP, X-Frame-Options, etc.)
- Comprehensive error handling
- Structured logging

What we learned

Technical Skills

Full-stack development
AI integration and prompt engineering
Production hardening (security, rate limiting, error handling)
Database migrations, schema design, and connection pooling
DevOps with Vercel and CI/CD

Design Principles

Safety-first design
User-centered development
Progressive enhancement
Comprehensive automated testing

AI Development

Prompt engineering dramatically affects output
Always validate AI output
Fall-back strategies are essential
Crisis detection is difficult but critical

Soft Skills

Breaking down complexity
Writing documentation
Learning with an AI pair programmer

What's next for Inner Harbor

Short-term enhancements

User authentication
Mobile app (iOS/Android)
Ritual customization
Anonymous ritual sharing

Medium-term features

ML-based personalized recommendations
Voice-guided rituals
Wearable integrations for auto-detected stress
Multilingual support

Long-term vision

Clinical validation
Therapist dashboards
Community ritual libraries
Research platform using anonymized data

Immediate next steps

Gather user feedback
Improve AI personalization
Add analytics for ritual effectiveness
Explore partnerships with mental health organizations

Inner Harbor represents my vision for the future of accessible mental wellness—AI-powered, personalized, evidence-based, and always available when you need it most.

Live Demo: https://inner-harbor-eta.vercel.app
GitHub: https://github.com/visshbala21/innerHarbor