Anjola-jolly/cultural-design-ai
Type de record:
software
Créateur:
Anj
Hôte:
Agentic AI framework for culturally grounded graphic design generation, focused on Yoruba Nigerian visual principles. Built with RAG, ResNet-50 CNN classifier, and DALL-E 3 via N8N.
# Cultural Design AI 🎨
An agentic AI framework for culturally grounded graphic design generation, with primary focus on Yoruba Nigerian visual principles and comparative analysis of Bauhaus and Swiss International Style traditions.
---
## Overview
Mainstream AI generative tools consistently misrepresent African visual traditions, producing stereotypical or culturally inauthentic outputs. This project addresses that gap by building a retrieval-augmented generation (RAG) pipeline that queries a structured cultural knowledge base before generating design outputs, ensuring outputs are grounded in real cultural principles rather than generalized AI training data.
Built as part of an MSc dissertation in Artificial Intelligence Technology at Northumbria University.
---
## How It Works
```
Design Prompt
↓
AI Agent (GPT-4o)
↓
Vector Store Retriever queries a 33-entry cultural taxonomy
↓
Culturally Enriched Prompt
↓
DALL-E 3 Image Generation
↓
Culturally Grounded Design Output
```
---
## Key Components
### 1. Cultural Knowledge Taxonomy
A 33-entry structured JSON knowledge base covering three design traditions:
| Tradition | Entries | Coverage |
|-----------|---------|----------|
| Yoruba (Nigerian) | 25 | Colour symbolism, Adire textile patterns, geometric vocabulary, Orisha associations, philosophical principles |
| Bauhaus | 4 | Primary colour theory, geometric abstraction, form-function philosophy |
| Swiss International Style | 4 | Grid systems, Helvetica typography, objective communication |
Each entry encodes: tradition · category · element · visual description · cultural meaning · philosophical context · design application · AI distinction notes.
### 2. CNN Classifier (ResNet-50)
- **Model:** ResNet-50 with ImageNet pre-training
- **Dataset:** 240 culturally annotated images
- **Classes:** Yoruba · Bauhaus · Swiss International Style
- **Validation Accuracy:** 91.67%
- **Training:** Two-stage transfer learning (frozen backbone → full fine-tuning)
- **Aug …