Brandmind: Building a multi-agentic framework for brand identity

DS-UA 301 Spring 2026
Sylvia Zhang · Helen Wei · Fiona Kao

---

Overview

Brandmind is a multi-agent LLM system that takes a natural language brand description and outputs a complete brand identity starter kit, including font pairings, color palette, and brand tone/voice — with design reasoning. Built with LangGraph, GPT-4o, and structured function calling against real typography and color datasets.

---

System Architecture

Brand Brief (text + optional image)
        │
        ▼
┌─────────────────┐
│  Planner Agent  │  → classifies brand archetype, extracts constraints
└────────┬────────┘
         │
         ▼
┌──────────────────────┐
│  Generator Agent     │  → calls font_lookup(), color_retrieve(), heuristic_search()
└────────┬─────────────┘
         │
         ▼
┌─────────────────┐       pass → Brand Kit Output
│   QC Agent      │  ──▶
└────────┬────────┘       fail → revision feedback → Generator (max 3 iterations)
         │
         ▼
   Streamlit Frontend

---

Repo Structure

brandmind/
├── state.py                  # shared LangGraph state schema — do not edit alone
├── agent1_planner.py         # Planner Agent (Fiona)
├── agent2_generator.py       # Design Generator Agent (Sylvia)
├── agent3_qc.py              # QC Agent (Helen)
├── graph.py                  # full LangGraph pipeline (assembled together)
├── app.py                    # Streamlit frontend
├── tools/
│   ├── font_lookup.py
│   ├── color_retrieve.py
│   ├── heuristic_search.py
│   └── wcag_check.py
├── data/
│   └── emotion_labeled_palettes.csv
├── requirements.txt
├── .env.example
└── README.md

---

Quickstart

1. clone and install

git clone https://github.com/your-org/brandmind.git
cd brandmind
pip install -r requirements.txt

2. Set up environment variables

cp .env.example .env
# fill in your keys

.env.example:

OPENAI_API_KEY=your-key-here
GOOGLE_FONTS_KEY=your-key-here   # optional, falls back to curated list

3. run individual agents (for development)

python agent1_planner.py
python agent2_generator.py
python agent3_qc.py

4. run the full pipeline

python graph.py

5. launch the Streamlit app

streamlit run app.py

---

running on Google Colab

!git clone https://github.com/your-org/brandmind.git
%cd brandmind
!pip install -r requirements.txt -q

import os
os.environ["OPENAI_API_KEY"] = "sk-..."

Upload emotion_labeled_palettes.csv to the Colab session, or mount Google Drive.

---

Agents

Agent	File	Owner	Description
Planner	agent1_planner.py	Fiona	Classifies brand archetype, extracts design constraints
Generator	agent2_generator.py	Sylvia	Retrieves fonts + colors + rules, assembles draft brand kit
QC	agent3_qc.py	Helen	Checks WCAG contrast, archetype coherence, constraint satisfaction

---

datasets

Dataset	Source	Used For
Google Fonts	fonts.google.com / API	Font candidates for font_lookup()
Emotion-Labeled Color Palettes	Kaggle	Palette retrieval for color_retrieve()
EmoSet	EmoSet	Visual emotion grounding

---

baselines

System	Description
BrandMind (ours)	3-agent pipeline + tools + shared memory + self-correction
Baseline 1: Zero-shot GPT-4o	Single LLM call, no tools, no agents
Baseline 2: RAG only	Retrieval + single LLM pass, no revision loop
Baseline 3: Fontjoy	Rule-based font pairing only, no color or tone

---

evaluation metrics

• Constraint Satisfaction Rate — % of outputs honoring stated brand constraints
• Human Preference Score — Likert 1–5, 20 participants, blind A/B vs. Baseline 1
• WCAG Pass Rate — % of palettes passing WCAG 2.1 AA contrast (programmatic)
• Archetype Coherence Score — GPT-4o judge, 1–5 scale

---

due dates

Date	Milestone
Mar 8	Milestone 1 — Proposal
Apr 5	Milestone 2 — Individual agents working, baseline comparisons
May 3	Milestone 3 — Full pipeline, evaluation, ablation study
TBD	Final — Live presentation + GitHub submission

---

related Work

• Choi & Hyun (2024). Typeface network and the principle of font pairing. Scientific Reports.
• Wu et al. (2023). AutoGen: Enabling next-gen LLM applications via multi-agent conversation. arXiv:2308.08155.
• Madaan et al. (2023). Self-Refine: Iterative refinement with self-feedback. arXiv:2303.17651.
• Bahng et al. (2018). Coloring with words. ECCV.