HackStack

Project Info

Winner

Google: Best Use of Gemini; Nobel Era Group: Best Projects Leveraging LLM Agents Supported by Nobel Era Group & Berkeley RDI

DreamDirector

Devpost

Inspiration

We love immersive experiences. We dreamed of a platform where you could not just imagine the story, but see it unfold like a movie with adaptive music that responds to your emotions. The latest advances in multimodal AI from Google, including Gemini, Imagen 3.0, Veo 2.0, and Lyria RealTime, gave us the tools to finally build it: a true AI storyteller that weaves narrative, visuals, and dynamic audio together in real time.

What it does

DreamDirector is your personal AI Cinematic Director. The entire experience starts with you; you can type in a prompt for any world you want to build, any story you want to tell. Whether it's a 'cyberpunk mystery' or a 'fantasy quest', our multi-agent AI system collaborates to bring that world to life. The system generates stunning 4K images with visual consistency tracking, produces cinematic 8-second video sequences for dramatic moments, and creates adaptive music that evolves with your story's emotional tone. Every choice you make matters, with the AI maintaining persistent memory of characters, locations, and world state across your entire journey.

How we built it

Architecture & Technology Stack Frontend (React 18 + Modern Web) React 18 with Vite for lightning-fast development Tailwind CSS with custom cinematic themes Framer Motion for advanced animations and transitions React Router Dom for seamless navigation Axios for API communication Backend (FastAPI + Python) FastAPI with async/await for high-performance API handling Uvicorn ASGI server with real-time capabilities Pydantic for data validation and type safety Python-multipart for media file handling CORS middleware for cross-origin support AI & Multi-Agent System (Google ADK + GenAI) Google ADK (Application Development Kit) for agent orchestration Google GenAI SDK with Gemini models for intelligent reasoning Google Imagen 3.0 for consistent visual generation Google Veo 2.0 for cinematic video sequences Google Lyria RealTime for streaming adaptive music Multi-Agent Workflow We architected DreamDirector as a team of specialized AI agents working through the Google ADK framework: Story Director Agent Orchestrates narrative flow using Gemini's reasoning capabilities Manages three-act story structure and pacing Generates meaningful user choices with preview consequences Maintains character development and plot consistency Visual Consistency Agent Uses Imagen 3.0 with character/location reference tracking Maintains visual DNA across all generated content Ensures professional cinematographic composition Applies consistent artistic style and color palettes Adaptive Music Composer Leverages Lyria RealTime for streaming audio generation Creates layered compositions (atmospheric, character, action themes) Responds to emotional tone and story tension in real-time Supports multiple genres from cyberpunk synthwave to orchestral fantasy Media Orchestrator Coordinates multi-modal generation timing Optimizes resource usage across Google's APIs Handles video generation with image seeds for consistency Manages media balance for optimal user experience User Choice System Our choice system goes beyond simple branching narratives: Dynamic Choice Generation where AI creates contextual choices based on current story state Visual Preview System where users see outcomes through generated imagery Persistent Consequences where every decision affects character relationships and world state Adaptive Difficulty where story complexity adjusts to user engagement patterns Memory Integration where choices influence future scene generation and character interactions

Challenges we ran into

Orchestrating multiple cutting-edge AI models in real-time presented significant challenges. Ensuring visual consistency across Imagen 3.0 generations required sophisticated prompt engineering and reference image tracking. Integrating Lyria RealTime's streaming audio with FastAPI's async architecture demanded careful threading and session management. The most complex challenge was building a coherent multi-agent system where each AI agent could communicate through structured function calls while maintaining narrative coherence and fluid user experience.

Accomplishments we're proud of

We successfully created the first truly multimodal storytelling engine that combines Gemini's reasoning, Imagen's visuals, Veo's cinematography, and Lyria's music into a seamless experience. Our multi-agent architecture represents a breakthrough in AI collaboration; each agent specializes in its domain while contributing to a unified creative vision. We're especially proud of our visual consistency system that maintains character appearance and world continuity across an entire story session, and our adaptive music system that creates Hollywood-quality soundtracks in real-time.

What we learned

This hackathon was a masterclass in practical multimodal AI engineering. We learned how to architect complex agent systems using Google ADK's tool framework, mastered advanced prompt engineering for consistent visual generation, and discovered the nuances of real-time audio streaming. Most importantly, we proved that sophisticated AI agent collaboration can create experiences far richer than any single model can do alone.

What's next

We envision expanding DreamDirector into a comprehensive creative platform. Our roadmap includes: Multiplayer Adventures with shared AI-crafted worlds for collaborative storytelling Creator Tools to allow users to define custom genres and fine-tune agent behaviors Extended Media Support for integration with more Google AI APIs as they become available Mobile Experience for native iOS/Android apps with offline story caching Community Features for story sharing, remixing, and collaborative world-building Submission for the UC Berkeley AI Hackathon 2025 - Creativity Track DreamDirector represents the future of AI-powered interactive entertainment, showcasing how multiple specialized agents can collaborate to create magical user experiences. Problem and Market Opportunity The creator economy demands high-quality multimedia content, but individual creators face immense time and cost barriers. DreamDirector democratizes cinematic storytelling by turning weeks of traditional production into seconds of AI-powered generation, making professional-quality interactive media accessible to everyone. Multi-Agent Innovation and Technical Excellence Our core innovation is a sophisticated multi-agent system built on Google ADK that goes far beyond simple prompt chaining. Each agent maintains specialized knowledge domains and communicates through structured function calls with persistent state management. This architecture enables complex creative workflows while maintaining reliability and coherence. Scalability and Impact Built on Google's cloud-native APIs with async FastAPI architecture, DreamDirector is designed for massive scale. Our vision is to keep the core storytelling experience free for creators, students, and hobbyists while building sustainable revenue through enterprise API licensing for professional studios seeking to integrate our multi-agent creative engine into their workflows. DreamDirector, in short, is the foundation for the next generation of interactive entertainment.

Analysis

Compare with all teams

View

Metric

Figures cover GitHub contributors during the hackathon window. A co-authored commit counts in full for each author, so per-member totals add up to more than the whole-team figures.

Technology

Found in codeClaimed only
  • CSSIn code
  • FastAPIIn code
  • HTMLIn code
  • JavaScriptIn code
  • PythonIn code
  • ReactIn code
  • StreamlitIn code
  • Tailwind CSSIn code

8 of 8 appear in the indexed code.

AI coding agents

No AI coding agent signals were found in this repository.

Detected from committed agent config files and commit authorship. Absence of a signal is not proof an agent was unused.

Codebase size

Source size

560 KB

Source files

40

Counts recognized source files only; vendored directories, binaries and lockfiles are excluded, so this is smaller than the repository on disk.

0 stars