DreamWeaver

AI bedtime companion that learns your family, speaks in your voice, and captures fleeting moments—transforming the nightly ritual into lasting magic.

Comment

💡 Inspiration

Every parent has felt it: the guilt of being too exhausted to read with enthusiasm, the frustration of finding age-appropriate content, the bittersweet knowledge that these tender moments are fleeting. We watched friends rush through bedtime stories after 12-hour workdays, saw grandparents miss daily moments because of distance, and witnessed beautiful conversations with children—gone forever because no one captured them.

DreamWeaver was born from a simple question: What if AI could strengthen family connection instead of replacing it?

When Gemini 3 was announced, we saw the opportunity to build something truly agentic—not just another LLM wrapper, but an AI companion that learns, anticipates, and acts autonomously to make bedtime magical.

🎯 What it does

DreamWeaver is an agentic AI bedtime companion that transforms the nightly ritual:

🎨 Story Engine (Gemini 3 Flash)

  • Generates personalized stories using child's name, age, and interests
  • Adapts themes based on weather, time of day, and recent conversations
  • Produces age-appropriate narratives with 4-layer safety validation

🎙️ Voice Cloning

  • Narrates stories in the parent's actual voice using voice cloning technology
  • 30-second recording creates unlimited personalized stories
  • Grandparents far away can still "read" bedtime stories

🌙 Sleep Sentinel

  • Real-time audio analysis detects sleep cues (breathing patterns, silence)
  • Automatically adjusts story pacing as child approaches sleep
  • Gracefully fades to ambient sounds when sleep detected (>85% confidence)

✨ Memory Curator (Gemini 3 Pro)

  • Automatically detects and preserves "golden moments"
    • Developmental milestones (first time saying a new word)
    • Emotional moments ("I love you, Daddy")
    • Curious questions ("Why is the sky blue?")
  • Captures child's actual voice for future playback
  • Weekly digest emails for parents

👵 Grandma Mode

  • One-tap secure sharing of moments with extended family
  • Magic links expire in 48 hours, 3-view limit
  • No account required for recipients
  • Privacy-first: not indexed, not public

🎙️ Live Mode (Gemini 2.5 Live)

Note: Gemini 3 Live not yet available, using Gemini 2.5 Live

  • Real-time conversational storytelling
  • Children can interrupt: "What's that star?" → AI answers in narrative
  • Voice direction: "Make her fly!" → Story adapts instantly
  • WebSocket streaming for <200ms latency

🏗️ How we built it

Architecture: Clean + Agentic

We implemented Clean Architecture with true multi-agent orchestration:┌─────────────────────────────────────────────────────────┐ │ BEDTIME CONDUCTOR (Meta-Agent) │ │ Goal: Child asleep + Memories captured + Joy maximized │ └─────────────────┬───────────────────────────────────────┘ │ ┌───────────┼───────────┬───────────┬──────────────┐ ▼ ▼ ▼ ▼ ▼ ┌──────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐ │ Story │ │ Voice │ │ Memory │ │ Sleep │ │ Safety │ │ Weaver │ │ Artisan │ │ Curator │ │Sentinel │ │Guardian │ │(Gemini 3)│ │ (TTS) │ │(Gemini 3│ │(Audio │ │(4-layer)│ │ Flash │ │ │ │ Pro) │ │Analysis)│ │ │ └──────────┘ └─────────┘ └─────────┘ └─────────┘ └─────────┘

Tech Stack

Frontend:

  • React 19 + TypeScript + Vite
  • TanStack Query (server state) + Zustand (client state)
  • Tailwind CSS + Radix UI (accessible primitives)
  • Service Workers (PWA offline capability)

Backend (Hono + Node.js ESM):

  • Clean Architecture (Domain → Application → Infrastructure)
  • Supabase (PostgreSQL + pgvector for RAG)
  • Gemini 3 Flash (fast story generation)
  • Gemini 3 Pro (complex reasoning for golden moments)
  • Gemini 2.5 Live (real-time voice interaction)
  • Google Cloud TTS with voice cloning
  • Upstash Redis (session state + caching)

Infrastructure:

  • Vercel (App + API hosting)
  • Cloudflare Workers (WebSocket relay for Live Mode)
  • Supabase Realtime (event streaming)

Gemini 3 Integration

We use three Gemini models strategically:

  1. Gemini 3 Flash (gemini-3-flash-preview): Fast story generation

    • Cost-effective ($0.04/story avg)
    • Structured JSON output with Zod schemas
    • Personalization via family context embeddings

  2. Gemini 3 Pro (gemini-3-pro-preview): Complex reasoning

    • Golden moment detection (milestone classification)
    • Multi-turn conversation synthesis
    • Safety verification (fail-closed design)

  3. Gemini 2.5 Live (gemini-live-2.5-flash-native-audio): Real-time voice

    • Bidirectional audio streaming (WebSocket)
    • <200ms latency via edge processing
    • Tool calling for dynamic story elements

Proof of Gemini 3 usage: GET /api/v1/meta/gemini-models returns resolved model names.

🧠 What makes it "Agentic"?

DreamWeaver isn't an LLM wrapper—it's a goal-directed multi-agent system:

Characteristic LLM Wrapper DreamWeaver (Agentic)
Trigger User command Context + Autonomous
Memory Session-only Persistent family context
Decision-making None (execute prompt) Goal-directed with trade-offs
Learning None Active preference adaptation
Proactivity Never Suggests, warns, anticipates

Example: Autonomous Sleep Detection

The Sleep Sentinel agent:

  1. Analyzes audio stream in real-time (RMS + FFT)
  2. Detects breathing cadence patterns
  3. Publishes SLEEP_CUE events to message bus
  4. Story Weaver autonomously responds by:
    • Slowing narrative pacing
    • Simplifying sentence structure
    • Preparing gentle conclusion
  5. At 90% confidence → graceful fade to ambient sounds

No user intervention required. The agents coordinate via event bus to achieve the shared goal: child peacefully asleep.

The Bedtime Conductor

Our meta-agent manages a goal stack (LIFO, max depth 5): ```typescriptinterface Goal { type: 'STORY_COMPLETED' | 'CHILD_ASLEEP' | 'MEMORY_CAPTURED' priority: number constraints: Constraint[] }

Example goal resolution:

  1. Goal: Child asleep in 15 minutes
  2. Constraint: Capture at least 1 golden moment
  3. Trade-off: Story length vs. interaction depth
  4. Decision: Gemini Pro analyzes context → suggests 8-minute story with 2 interaction points
  5. Execution: Agents autonomously coordinate
  6. Adaptation: If child still awake at 12 min, switch to "gentle lullaby" mode

🚧 Challenges we ran into

1. Real-time Audio Latency

Problem: Initial Gemini Live integration had 2-3 second latency, breaking immersion.

Solution:

  • Moved WebSocket relay to Cloudflare Workers (edge processing)
  • Implemented binary PCM16 streaming (no base64 overhead)
  • Result: <200ms end-to-end latency

2. Voice Cloning Quality

Problem: Early voice clones sounded robotic ("uncanny valley").

Solution:

  • Switched to Google Cloud TTS Chirp 3 Instant Custom Voice
  • Implemented 30-second recording with prompt guidance
  • Added fallback to high-quality preset voices if cloning fails
  • Result: 85% user acceptance rate in testing

3. Safety at Scale

Problem: AI-generated content for children requires zero-tolerance for harmful content.

Solution: 4-layer defense system:

  1. Prompt-injection defense (regex patterns)
  2. Content policy patterns (fast blocklist)
  3. AI semantic check (Gemini Pro verification, fail-closed)
  4. Heuristic validation (age-appropriateness checks)

All content passes through SafetyGuardian before reaching children.

4. COPPA Compliance

Problem: Children under 13 = strict privacy requirements.

Solution:

  • Parent creates account, adds children (no direct child registration)
  • Row-Level Security (RLS) in Supabase ensures data isolation
  • No behavioral advertising, no cross-site tracking
  • Full data export + deletion within 24 hours
  • All voice recordings user-controlled (opt-in)

5. Cost Control

Problem: Gemini API costs could spiral with unlimited story generation.

Solution:

  • Strategic model selection (Flash for 90% of operations)
  • Aggressive caching (Upstash Redis + browser cache)
  • Free tier: 5 stories/week (prevents abuse)
  • Per-session cost guardrails (max $0.50/session)
  • Result: $0.04 average cost per story

🎓 What we learned

Technical Insights

  1. Gemini 3's structured outputs are game-changing

    • Direct JSON schema validation eliminates parsing errors
    • 10x reduction in prompt engineering complexity

  2. Voice cloning is production-ready

    • 30-second samples produce convincing results
    • Parents cried during testing (emotional, not technical failure!)

  3. Edge computing matters for real-time AI

    • Cloudflare Workers + Gemini Live = <200ms latency
    • Traditional server-side would be 2-3x slower

  4. Clean Architecture scales beautifully

    • Added Live Mode without touching domain layer
    • Swapped TTS providers in 20 lines of code

Product Insights

  1. Parents want AI to amplify them, not replace them

    • Voice cloning = most loved feature
    • Automation without authenticity = rejected

  2. Golden moments are the moat

    • Users tolerate bugs if memories are preserved
    • Memory vault = 90% retention driver in beta

  3. Grandma Mode is viral

    • 40% of beta users shared a moment within 7 days
    • Emotional sharing > traditional referrals

🚀 What's next for DreamWeaver

Immediate (Post-Hackathon)

  • [ ] Gemini 3 Live integration (when available)
  • [ ] Multi-language support (starting with Spanish)
  • [ ] Native iOS/Android apps (currently PWA)
  • [ ] Dream Companions gamification (unlock characters after N stories)

3-Month Roadmap

  • [ ] Smart home integrations (dim lights when sleep detected)
  • [ ] Weekly Time Capsule emails (automated highlights)
  • [ ] Voice marketplace (licensed celebrity voices)
  • [ ] Multi-child sessions (siblings share story with arbitration)

Long-term Vision

  • [ ] Physical "Dream Companion" hardware (smart speaker for kids)
  • [ ] B2B partnerships (pediatricians, daycares, schools)
  • [ ] International expansion (40M parents in US → 200M globally)
  • [ ] AI therapy integration (address bedtime anxiety, nightmares)

🏆 Why DreamWeaver wins

1. True Agentic AI

Not an LLM wrapper. Multi-agent orchestration with autonomous decision-making.

2. Deep Gemini 3 Integration

Strategic use of Flash (speed), Pro (reasoning), and Live (real-time).

3. Emotional Moat

Voice cloning + memory vault = emotional switching cost.

4. Production-Ready

COPPA compliant, 4-layer safety, real users in beta.

5. Business Model

$9.99/month Premium tier, viral growth via Grandma Mode, $500M TAM.

🔗 Links

🙏 Acknowledgments

Built with ❤️ using:

  • Google Gemini 3 (Flash, Pro) + Gemini 2.5 Live
  • Supabase (database + auth)
  • Vercel (hosting)
  • Cloudflare Workers (WebSocket relay)
  • Google Cloud TTS (voice cloning)

Special thanks to the Gemini team for building APIs that make true agentic AI possible.

DreamWeaver: Your voice. Their stories. Memories forever. 🌙✨

Built With

Share this project:

Updates