Cubtales - AI Storybook Platform
An AI-powered storybook generator where parents create custom characters, personalities and worlds - and the platform writes and illustrates a fully consistent, age-appropriate book they can read together.
Overview
Cubtales started with a simple question: what if a child who didn't want to read could make their own book? Parents create characters - how they look, their personality, their heroes and villains - and the platform generates a fully illustrated story that maintains consistency across every page and even into follow-up series. This was my first major agentic AI project, coordinating multiple models for narrative, illustration and safety. Built end-to-end as a solo project, with phase 1 targeting release in 2026.
The Inspiration
A friend's child was learning to read but had no interest in the books teachers could offer. The question was simple: what if he could make his own? Cubtales was born from that idea - a platform where parents and children create their own characters, worlds and stories, and AI does the heavy lifting of writing and illustrating a complete book they can enjoy together.
Character Creation & Memory
Parents start by creating characters - defining how they look, what they wear, their personality, their heroes (a loyal fluffy dog) and their villains (the evil Mister Caterpillar, or whatever they dream up). The platform stores these details as persistent character profiles. Throughout the book, and across follow-up series, the guardrails ensure consistency: a red shirt stays red, a fluffy dog stays fluffy, personality traits carry through into dialogue and actions. This isn't just prompt injection - it's structured character memory that feeds into every generation call.
Agentic AI Workflows
This was my first large-scale undertaking of agentic AI workflows. The generation pipeline coordinates multiple AI models across narrative writing, illustration generation and content safety. Cloud models (OpenAI and Claude) handle the primary generation, while local models via Ollama serve as a sandboxing layer for development and testing - reducing API costs significantly during iteration. The agentic team approach means each model has a focused role: one writes narrative, another validates age-appropriateness, another generates illustration prompts that maintain visual consistency with established character descriptions.
Content Safety & Guardrails
Every piece of generated content - text and images - passes through a multi-layer safety pipeline. Narrative guardrails ensure stories remain age-appropriate, avoiding violence, fear, inappropriate themes and anything that doesn't fit the tone parents have set. Illustration guardrails enforce visual consistency and prevent the generation of unsuitable imagery. The system avoids copyrighted art styles and maintains ethical AI usage throughout. In testing, the moderation pipeline has produced zero unsafe outputs.
Reading & Distribution
Parents can read the finished book directly in the app, with a page-turning reading experience designed for parent-and-child sessions. Books can also be downloaded as PDFs for offline reading or printing at home. A future feature will allow parents to order hard or softback printed copies of their child's stories - turning AI-generated content into a physical keepsake.
Engineering
Built end-to-end as a solo project on the latest Laravel and Nuxt stack. The backend handles character storage, generation orchestration, PDF compilation and content moderation. Pest tests cover the generation pipeline and guardrail logic, with Cypress end-to-end tests covering the parent-facing creation and reading flows. The project is still a work in progress, with phase 1 targeting release in 2026.
Technology Stack
Technical Challenges
Visual Consistency Across Generated Illustrations
Built a structured character memory system that feeds appearance details, clothing, colours and physical traits into every illustration prompt. The system maintains visual consistency not just within a single book, but across follow-up series - ensuring characters are always recognisable regardless of scene or narrative context.
Age-Appropriate Content Generation at Scale
Implemented a multi-layer guardrail pipeline that validates both narrative text and generated images for age-appropriateness. Each generation passes through safety checks before reaching the user, with the system tuned to reject content that doesn't match the parent's chosen tone and the child's age range.
Cost-Effective AI Development With Local Models
Set up Ollama locally to run open-source models for development and sandboxing, dramatically reducing API costs during the iteration-heavy prompt engineering and workflow design phase. Cloud models are reserved for production generation, while local models handle testing, experimentation and agentic team coordination.