Case Study: AI-Driven Lay Flat Generation

Overview

The AI-Driven Lay Flat Generation project explored how we could use computer vision and AI to automate the creation of clean, on-brand product imagery — reducing manual photo editing and improving merchandising scalability. I led early product development, defining the system architecture, workflow logic, and quality-control mechanisms that set the foundation for future AI-powered merchandising tools.

Timeline: 4 months (R&D → MVP Design)
Impact:

• Designed an end-to-end visual AI workflow architecture
• Established repeatable design patterns for AI evaluation and fallback logic
• Informed subsequent automation features like background removal and moderation tooling

The Problem

Retail clients needed consistent, high-quality “lay flat” images (garments displayed neatly without a model). Producing them manually was slow, expensive, and inconsistent — especially when scaling across thousands of SKUs and brands.

Our goal was to reduce this dependency on manual editing while maintaining brand accuracy and aesthetic control.

My Approach

Discovery

I began by mapping existing asset ingestion and classification workflows to identify where automation could deliver the most impact. Working closely with Engineering and Design, I defined the required metadata, tagging, and AI integration points that would allow a model to:

• Detect and classify image types (on-model, swatch, lay flat, etc.)
• Rank images based on resolution, transparency, and background quality
• Generate new lay flat images where gaps existed

Design

I architected a modular system using Langsmith that blended automation with human oversight:

• Image Ranking Logic: Selected the best visual input based on quality signals
• Conditional Logic: Applied product-specific rules (e.g., skip lay flat generation for home goods)
• AI Evaluators: Scored outputs on pose accuracy, clarity, and brand consistency
• Human-in-the-Loop Interface: Provided moderation controls and fallback paths for edge cases (e.g., graphic tees or non-standard silhouettes)

Delivery

To validate feasibility, I scoped an MVP that could process small image sets through checkpoint-based workflows — ensuring traceability and easy debugging. I built a lightweight web interface using Claude to visualize prompt results directly from LangSmith. This internal tool allowed the team to quickly review outputs, spot inconsistencies, and iterate on model behavior in real time — significantly improving testing speed and prompt engineering. While the project remained experimental during my time, it served as a valuable learning layer that refined our approach to AI evaluation and visual asset generation.

Outcome

While the system remained in MVP phase, it achieved several key outcomes:

• Established a repeatable framework for AI image generation and evaluation
• Informed future automation tooling (background removal, moderation, metadata tagging)
• Helped the team define AI ethics and fallback principles for visual workflows

This project served as a test bed for scaling visual AI capabilities responsibly — bridging experimentation and production readiness.

Reflection

This project taught me that successful AI integration in product management isn’t just about accuracy — it’s about orchestration. Balancing automation with human judgment, performance with governance, and innovation with reliability is where the real product thinking happens. It reminded me that AI projects thrive when they’re not just technically impressive, but operationally sound.