🎭 ShakeScript - AI-Based Story Creation Platform

ShakeScript is a cutting-edge, AI-powered storytelling system designed to generate immersive, multi-episode narratives with rich characters, evolving plots, and long-term memory. By blending the strengths of GPT-4o and Google Gemini, it overcomes traditional limitations in AI storytelling—like token constraints and inconsistent narratives.

---

🔍 Problem Statement

Traditional AI-generated stories often struggle with:

Challenge	Description
🔄 Narrative Coherence	Maintaining seamless connections between episodes
🧠 Token Limitations	Handling restricted context windows in large language models
👤 Character Consistency	Preserving character traits, relationships, and emotional states
📚 Extended Narratives	Structuring long stories into coherent, episodic chunks
💾 AI Memory Integration	Retaining relevant story context across episodes

---

🚀 ShakeScript: Your AI Storytelling Engine

ShakeScript enables long-form storytelling with AI memory, human feedback integration, and cultural nuance.

✅ Core Capabilities

• Accepts brief prompts (genre, trope, or plotline)
• Generates multi-episode stories with world-building and character arcs
• Maintains narrative continuity via metadata & embeddings
• Supports Hinglish storytelling
• Offers both AI-driven and human-in-the-loop episode refinement
• Uses a robust database + semantic embeddings for memory

---

📂 Architecture & Workflow

1️⃣ Prompt to Metadata Extraction (via Google Gemini)

• Endpoint: /stories
• Gemini extracts:
  • Characters: Names, roles, relationships, emotions
  • Settings: Detailed location descriptions
  • Structure: Exposition → Climax → Denouement
  • Theme & Tone: (e.g., Suspenseful, Romantic)
• Data stored in Supabase:
  • stories, characters tables

---

2️⃣ Episode Generation (via GPT-4o)

Initial Episode

• Uses structured metadata
• Generates the episode aligned to outline (e.g., Exposition)

Subsequent Episodes

• Retrieves up to 2-3 past episodes for context
• Embeddings fetch relevant content chunks for long-form continuity
• Ensures:
  • Character consistency
  • Thematic alignment
  • Narrative progression

Storage

• Saves episode content, title, summary, emotion in episodes table
• Splits episode into semantic chunks using SemanticSplitterNodeParser
• Vectorizes & stores in chunks table

---

3️⃣ Validation & Refinement (via Gemini)

AI Validation (validation.py)

• Checks:
  • Timeline alignment
  • Character location/motivation consistency
  • Dialogue and tone coherence
• Refines up to 3 times if inconsistencies found

Human Feedback Support (episodes.py)

• Users can refine via /refine-batch
• Gemini regenerates while preserving core elements

Batch Processing (refinement.py)

• Default batch size: 2 episodes
• Intermediate state stored in current_episodes_content

---

4️⃣ Memory Management

Supabase (db_service.py)

• Tracks:
  • current_episode
  • key_events, timeline
  • Character evolution

Embedding Service (embedding_service.py)

• HuggingFace embeddings vectorize story chunks
• Relevance scored based on:
  • Characters involved
  • Episode order
• Enables memory-aware story generation

---

5️⃣ API & Frontend Integration

FastAPI Backend

• Endpoints:
  • /stories – Create new story
  • /generate-batch – Batch generate episodes
  • /validate-batch – AI validation
  • /refine-batch – Human feedback and refinement
• Uses Pydantic models for structure (schemas.py)

Frontend (Planned)

• React/Next.js UI
• Features:
  • Episode display
  • Character profiles
  • Hinglish support
  • Real-time story updates

---

🛠️ Tech Stack

Category	Technologies Used
AI & NLP	GPT-4o, Google Gemini, HuggingFace Embeddings
Backend	FastAPI, Pydantic, Asyncio
Database	Supabase (PostgreSQL)
Embeddings & Retrieval	LlamaIndex (SemanticSplitterNodeParser), Supabase Vector DB
Language	Python 3.13 with type hints

---

🎯 Key Achievements

• 🏆 Multi-Episode Consistency – Maintains coherent, evolving narratives
• 💡 Token Limit Workaround – Smart retrieval with embeddings
• 👤 Character Evolution – Tracks traits, arcs, and relationships
• 🔁 AI + Human Refinement – Combines LLM polish with user feedback
• 🌍 Hinglish Support – Culturally tuned storytelling

---

🖼️ Visuals & Evaluations

🧩 Model Pipeline

---

📈 Radar Chart - Story Attribute Comparison

---

🪱 Worm Graph - Evaluation Metrics (out of 10)

---

🔮 Future Enhancements

Feature	Description
🎮 Interactive Storylines	Let users influence story direction via input parameters
🎧 TTS Narration	Audio playback support with Text-to-Speech
🧠 Custom AI Models	Fine-tune LLMs for specific genres or styles
📱 Frontend UI	Responsive, real-time React/Next.js interface
🔍 Smart Retrieval	Advanced hybrid/cosine similarity chunk search

---

💬 Final Thoughts

ShakeScript redefines AI-powered storytelling by:

• Solving token limitation challenges
• Supporting long-form, culturally nuanced storytelling
• Seamlessly blending LLMs, embeddings, and human input

🎉 Let the stories unfold — with ShakeScript, your narrative has no limits.

---