Amazon Health Services

Project: Lifestyle Image Tool (LIT)
Roles: Lead Design Technologist
Year: 2025-2026

Description: Strategy, concept, design, and development of an AI application for the generation and modification of lifestyle images used in large-scale healthcare marketing.

What's the Problem Being Solved?

Amazon Health Services (AHS) marketing managers needed to run high-volume, data-driven marketing experiments, but a two-to-four week creative production cycle made that pace impossible, blocking the team's ability to test and learn at the speed marketing personalization requires.

• Bottleneck: Creative brief, production, and review took two to four weeks per cycle, far too slow for high-volume experimentation.

• What was needed: A creative, intuitive solution enabling marketing managers themselves to produce and modify high-quality lifestyle images.

• Immediate goal: A 25x increase in monthly creative testing capacity, mapped to 200+ images and 50+ videos produced per month.

• Long-term vision: Scale the production tooling to support hyper-personalized marketing.

The strategy behind LIT pairs every gain in production speed and testing scale with automated quality assessment, so growth in volume never comes at the cost of brand trust. My key assumption was that optimizing a manual human process led by a small creative team today would build the foundation to expand toward the agentic, hyper-personalized marketing scale of the near future.

Who Are We Solving It For?

The people using LIT day to day are Amazon Health marketing managers, each dropping into the tool with a specific creative task in mind rather than a desire to learn a new application. I mapped nine recurring jobs to be done to the four tabs of the product so each entry point matches a real request.

Jobs to Be Done

How We Solved It

I started by talking to end users and understanding their workflows. I considered leveraging existing design tools from Figma, Adobe, and others, but that approach required training, time, and security approvals, and was ultimately less cost-effective. Once stakeholders and I aligned on the strategy to achieve the business goals, I began designing and developing a working prototype directly, documenting workflows, acquiring use cases, and drafting engineering designs alongside the interface.

Why it matters: For creative image generation I look for variety in generative models, just like a brand team would look for variety in photographers. Some models respond to certain prompts better than others. By showing up to six frontier model responses I enable the type of choice expected in the creative production process.

Why it matters: All existing raw visual assets can be modified and repurposed as new content. By leveraging these minor changes and background replacements, I can exponentially scale our already approved lifestyle images for rapid experimentation.

Why it matters: Each generated image is evaluated by an LLM using a managed prompt with quality, brand and policy guidelines from a knowledge base as context. The response includes scores for each criterion, an overall score and short reasoning summary to help end users determine appropriate images. These scores also help accelerate the human review.

This scoring step depends on an LLM judge call, and before shipping it I tested whether adding step-by-step reasoning to that call was worth the added latency.

Reasoning didn't move the quality scores in this test, which meant the added latency wasn't buying better judgments for this use case, a small but concrete example of the kind of inference-latency tradeoff that only shows up once you're testing against a real model, not a mockup.

Example User Journey

Product & Prototype Architecture

Turning this from a prototype into a system engineered for scale meant three architecture decisions did the heavy lifting. Parallel multi-model generation and fire-and-forget saves keep the multi-model compare-and-score experience fast and non-blocking, so marketers see results as each model finishes instead of waiting on the slowest. Hash-based URL routing makes that state shareable and deep-linkable within Harmony's platform constraints, turning individual sessions into reusable, trustworthy references. An automatic evaluation agent moves scoring off the critical path onto an event-driven trigger, an architecture decision that scopes the product beyond today's human reviewers toward a future where agentic users can generate, evaluate, and act on images with no person required in the loop.

The table below represents the state architecture of the front-end and relevant back-end services, a distinction that determines what can fail safely, what needs to persist, and where session data actually lives.

Development Methodology

Over the course of LIT development the available coding tools and agents have drastically changed. The quirky logos remain the same, as does the development methodology, which includes grounding the agents in UX best practices and consistent patterns within the design system. From there I use highly detailed UX prompts with lo-fi mock-ups like the example below as pattern references. One coding agent helps me develop front-end, another back-end (although I handle most of the cloud infrastructure), and a third checks their work.

Design System & Interaction Design

Because AI coding agents made prototyping and iteration fast, most of my energy went into the differentiating, frontier-model-driven features rather than the interface chrome around them. LIT is an internal application that required real design pattern flexibility with all the other bells and whistles baked in. Amazon design systems are purpose built for retail applications or the cloud console, so Material Design 3 was an easy choice.

Because prompting carries the highest cognitive load, the interaction design is built around keeping that load focused on the creative work rather than the interface around it.

Where possible, use cases that normally require advanced prompt engineering were replaced with intuitive configuration UI, and prompting tips were provided in context.

What We Launched

High Level Scope

Execution Examples

Image Generation v1.6 - Async inference of one prompt to six models

Image Generation v1.12 - Async inference of one prompt to three top ranked models

Image Generation v1.12 - Automated image review and reasoning

Variant Generation v1.12 - Digital Asset Manager Search Integration

Image Editing v1.9 - Graphic overlay feature

Variant Generation v1.12 - Prompt abstraction for background color replacement

Video Generation v1.12 - Prompt and configuration deep-linking

Business Impact

From the initial prototype in late 2025 to the heavily iterated product of 2026, LIT has been used to generate over 4,000 images and videos. The approval rate of these assets started at 1.2% and stands at 22.4% today. Approved assets have been used in 311 marketing experiments viewed by over 1.6 million Amazon Health customers. While performance marketing metrics are confidential, the use of LIT assets has resulted in $1.8m in savings on outside agency and photography services at an average cost of $0.78 per asset. In May 2026 LIT was 254% above goal for generated images and 78% of goal for video. In June 2026 LIT was 342% of goal for images and 344% of goal for video.