Gen AI Bannerization Platform

Project Overview

Ever wondered what happens when you throw LLMs, Vision Transformers, and distributed computing at a design team's nightmare? You get a system that went from 6 hours per banner to 1 minute — and I got to build it.

Gen AI Bannerization is an end-to-end intelligent automation platform that revolutionized how Bajaj Finserv creates marketing banners. We didn't just automate design — we built an AI assembly line with 98% quality accuracy that produces 9,000+ unique banners daily.

The Challenge

Picture this: A design team drowning in repetitive banner requests — thousands of variants for web, mobile, email across 20+ product lines. Manual creation was the bottleneck; campaigns waited weeks to go live. We needed to make the impossible possible.

Key Features

• Multi-Agent Design System: Built an orchestra of specialized AI agents (Strategist, Background Designer, Foreground Designer, Developer) using LangGraph that collaborate to generate banners from simple campaign briefs

• Vision Transformer QC: Fine-tuned ViT model achieving 98% accuracy in detecting design defects, brand violations, text overflow, and color compliance issues — with per-banner confidence scoring

• Parallel Processing Engine: Architected RabbitMQ + Celery distributed system handling 100+ simultaneous generation tasks without breaking a sweat

• U²-Net Image Segmentation: Custom fine-tuned deep segmentation network for precise background removal with 94%+ accuracy on brand-specific imagery

• MetaPrompting Intelligence: Automatic extraction of high-value metadata for personalized content recommendations

Technical Deep Dive

The AI Agent Architecture

Campaign Brief → Strategist Agent → Background Designer → U²-Net Segmentation 
→ Foreground Designer → Developer Agent → Banner Assembly → ViT QC → Publishing

Each agent is purpose-built: the Strategist analyzes brand guidelines and historical performance; the Background Designer wields diffusion models; the Developer Agent outputs production-ready SVG/HTML5.

Quality Control That Actually Works

Our ViT-based QC system went from 94% to 98% accuracy through continuous learning on custom failure cases. It catches:

• Text misalignment and overflow
• Logo placement violations
• Brand color deviations (>5% from spec)
• WCAG contrast compliance issues
• CTA rendering problems

Scale That Doesn't Break

Built on RabbitMQ + Celery with intelligent task routing:

• GPU-equipped workers handle image-heavy tasks
• Priority queues ensure QC never waits
• Kubernetes HPA auto-scales based on queue depth
• Dead letter queues catch and analyze failures

Results That Matter

Metric	Before	After	Impact
Banner Creation Time	6 hours	1 minute	99.7% ↓
Daily Production	~50	9,000+	180x ↑
Manual Design Effort	100%	30%	70% ↓
QC Accuracy	Human (~85%)	98%	15% ↑
Design Rework	Frequent	Rare	80% ↓

Tech Stack

AI/ML: Vision Transformers, U²-Net, Diffusion Models, LangGraph, GPT-4 Backend: Python, FastAPI, Flask, Celery Infrastructure: RabbitMQ, Redis, Docker, Kubernetes Computer Vision: OpenCV, ONNX Runtime, OpenVINO Integration: Adobe Experience Manager (AEM), Pimcore

What I Learned

Building AI systems at scale is 20% model training and 80% systems engineering. The hardest part wasn't getting the models to work — it was making them work reliably at 9,000 requests per day while maintaining quality guardrails that real stakeholders trust.

---

This project taught me that the best AI systems are the ones users forget are AI — they just work, at scale, reliably.