Unwind

main page
login page
dashboard
voice inbox
personal vaults
categories

Inspiration

We've all experienced it - lying awake at 3 AM with our minds racing through tomorrow's tasks, last week's worries, and random ideas. The more we try to organize these thoughts, the more overwhelming it becomes. Traditional to-do apps demand structure when our minds are anything but structured.

What if you could just... speak?

That's where Unwind was born. We wanted to build something that meets people where they are - in the chaos - and transforms mental noise into clarity without forcing them to context-switch into "productivity mode."

What it does

Unwind is an AI-powered voice-first mental health and productivity app that:

Listens Without Judgment: Speak freely to an empathetic AI agent powered by ElevenLabs - no interruptions, no structure required
Understands Context: AI recognizes not just what you said, but how much you care about it through emotional weight analysis
Auto-Organizes Everything: Transcripts are processed by Claude 3.5 Sonnet to extract tasks, ideas, worries, and errands
Smart Categorization: Items are automatically sorted into vaults (Tasks, Ideas, Worries, Health, Relationships, Errands, Recurring)
Intelligent Prioritization: Uses a custom scoring algorithm combining urgency, importance, and emotional weight:

$$\text{Priority Score} = 0.4 \times \text{Urgency} + 0.3 \times \text{Importance} + 0.3 \times \text{Emotional Weight}$$

Production-Ready Monitoring: Full observability with Sentry tracking AI performance, token costs, and user sessions

How we built it

Tech Stack

Frontend: Next.js 16 + React 19 with server components
Voice AI: ElevenLabs conversational agent with custom prompts
AI Processing: OpenRouter + Claude 3.5 Sonnet for task extraction
Database: Supabase (PostgreSQL + Auth + Real-time subscriptions)
Monitoring: Sentry with AI span monitoring and Session Replay
Deployment: Vercel with edge functions

Architecture Flow

User Voice → ElevenLabs AI Agent → Webhook → Supabase
                                         ↓
                                    Claude 3.5 (via OpenRouter)
                                         ↓
                                  Extract & Categorize Items
                                         ↓
                                  Calculate Priority Scores
                                         ↓
                                  Store in Database + Real-time Updates

Key Features Implemented

Real-time voice conversation with animated Orb visualizer
Webhook-based transcript processing
AI-powered item extraction with emotional weight scoring
Automatic categorization into 7 vaults
Interactive calendar and scheduling
Session replay and AI performance monitoring
Row-level security for user data

Challenges we ran into

1. Webhook Timing Issues

Problem: ElevenLabs sends transcripts via webhook after the call ends, but arrival time is unpredictable.

Solution: Implemented a frontend refresh mechanism that polls after 30 seconds, assuming webhook processing completes. Added comprehensive error handling for webhook failures.

2. Timezone Conversion Nightmare

Problem: Supabase stores UTC timestamps, but JavaScript parseISO() was interpreting them as local time, causing voice dumps to show 5 hours ahead.

Solution: Force UTC interpretation by appending 'Z' to timestamp strings:

const utcDateStr = dateStr.includes('Z') ? dateStr : dateStr.replace(' ', 'T') + 'Z'

3. AI Extraction Consistency

Problem: Claude would sometimes return invalid JSON or miss tasks hidden in emotional rambling.

Solution:

Implemented explicit JSON schema validation with Zod
Added retry logic with exponential backoff
Increased temperature for better categorization creativity
Fine-tuned prompts to handle emotional context

4. Sentry Session Replay Configuration

Problem: Sentry wizard installed SDK but Session Replay wasn't capturing sessions.

Solution: Manually added DSN to sentry.client.config.ts and boosted sampling rate to 100% for hackathon demo. Discovered the wizard only configured server-side monitoring by default.