undercurrent

record your unseen emotions.

Undercurrent is an emotional journaling web app that captures video recordings and uses multi-signal AI analysis to surface the emotions beneath the surface — including ones you may not have named yourself.

What it does

1. Record — webcam + mic captures your journal entry with live speech-to-text captions
2. Analyze — four AI signals run in parallel:
   • Gemini reads the transcript for semantic emotion content
   • Hume Voice reads prosody and tone from the audio
   • Hume Face reads facial expressions from a captured frame
3. Fuse — signals are weighted and blended into a single emotion profile
4. Surface — a gradient visualization, a "emotion beneath" insight, and a contradiction alert (when your words say one thing and your voice says another) are shown on the results page
5. Persist — entries are saved to MongoDB with the full signal breakdown, viewable later in the insights page and calendar

Stack

Layer	Tech
Framework	Next.js 16 (App Router)
Styling	Tailwind CSS
Auth	Clerk v7
Database	MongoDB Atlas + Mongoose
AI / NLP	Gemini 2.5 Flash
Voice emotion	Hume Expression Measurement API
Transcription	ElevenLabs Scribe v2
Video storage	Vercel Blob
Webhook verification	Svix

Project structure

src/
├── app/
│   ├── record/          # Webcam recording + live captions
│   ├── results/         # Live results while pipeline runs
│   ├── entries/         # Entry list + individual insights page
│   ├── calendar/        # Emotion calendar view
│   ├── calibrate/       # Onboarding / emotion color calibration
│   ├── colors/          # Color preference settings
│   ├── sign-in/         # Custom Clerk sign-in page
│   └── sign-up/         # Custom Clerk sign-up page
│   └── api/
│       ├── analyze/     # Core fusion pipeline (4 signals → fused emotions)
│       ├── transcribe/  # ElevenLabs speech-to-text
│       ├── upload/      # Vercel Blob audio upload
│       ├── entries/     # CRUD for journal entries
│       ├── user/        # Clerk → MongoDB user sync
│       └── webhooks/    # Clerk webhook handler
├── lib/
│   ├── fusion.ts        # fuseEmotions(), detectContradiction()
│   ├── gradients.ts     # computeGradientColors()
│   ├── recordingStore.ts # In-memory pipeline state (record → results)
│   ├── mongodb.ts       # Mongoose connection with hot-reload cache
│   └── models/          # User, Entry Mongoose schemas
└── components/
    └── NavBar.tsx

Getting started

1. Install dependencies

pnpm install

2. Set up environment variables

Create .env.local:

# Clerk
NEXT_PUBLIC_CLERK_PUBLISHABLE_KEY=
CLERK_SECRET_KEY=
NEXT_PUBLIC_CLERK_SIGN_IN_URL=/sign-in
NEXT_PUBLIC_CLERK_SIGN_UP_URL=/sign-up
NEXT_PUBLIC_CLERK_SIGN_IN_FORCE_REDIRECT_URL=/record
NEXT_PUBLIC_CLERK_SIGN_UP_FORCE_REDIRECT_URL=/calibrate
CLERK_WEBHOOK_SECRET=

# MongoDB
MONGODB_URI=

# Gemini
GEMINI_API_KEY=

# Hume
HUME_API_KEY=

# ElevenLabs
ELEVENLABS_API_KEY=

# Vercel Blob
BLOB_READ_WRITE_TOKEN=

3. Run locally

pnpm dev

Open http://localhost:3000.

Key design decisions

Multi-signal fusion — No single AI signal is reliable on its own. Gemini reads meaning but misses tone; Hume reads tone but misses context. The weighted blend (Gemini 55%, Hume voice 25%, Hume face 10%, Presage 10%) produces a more honest read than any one model alone.

Session ID guard — The results page subscribes to a module-level store. A session counter ensures that if a user hits "record more" quickly, the previous recording's in-flight pipeline can't overwrite the new recording's state.

Immediate navigation — The user is sent to /results before the pipeline finishes, so there's no waiting on the record page. The results page shows a loading state and updates live as signals complete.