Ecommerce ops · Production

Popsa uses Amazon Nova and retrieval-augmented generation to deliver personalised photo book title suggestions at scale

The problem

Most Popsa customers are not professional copywriters and defaulted to generic titles like 'France 2024' or even just 'Photos'. The existing rule-based Title Suggestion Graph algorithm relied on templates and could not generate creative, personalised suggestions.

Workflow diagram · grounded in source
1
User selects photos, app extracts metadata
trigger
“When users select photos for a Photo Book design, our mobile app reads metadata—such as timestamps and geocoordinates—from the images and runs on-device convolutional neural networks to extract relevant features.”
2
Geocoding and subject classification
ai_action
“it performs a reverse geocoding operation on any latitudes and longitudes included in the design, and then classifies the subject of the design based on object landmarks. This generates a description like "A skiing photobook with 21 phot…”
3
Retrieve few-shot examples from database
ai_action
“We created a database of example Photo Books and acceptable title suggestions. For a new Photo Book, we retrieved a few similar Photo Book designs and a random selection of their suggested titles.”
4
LLM generates title-subtitle-category triplets
ai_action
“Using Amazon Bedrock and Anthropic's Claude 3 Haiku, we seeded the conversation with these examples as <user> – <assistant> messages before appending the user's new design document as the final <user> message. This allowed the large lang…”
5
Stream parsing and validation
validation
“we extended the FastAPI to parse streams in real time, returning the first suggestion immediately upon validation. Additional suggestions continue streaming in the background, but the client already has something ready to display.”
6
Customer feedback loop
feedback_loop
“we relied on a feedback loop, where customers rated suggestions as positive, neutral, or negative. We also conducted multivariate testing with hundreds of thousands of users.”
Reported outcome

Switching to retrieval-augmented generative AI improved positive user feedback by 13% over the graph algorithm with Claude 3 Haiku, and Nova Pro further lifted satisfaction to 73% positive feedback.
Migrating to the ConverseStream API cut time to first suggestion by 35%, and the feature generated over 5.5 million personalised titles in 2025.

Reported metrics
Personalised titles generated in 2025over 5.5 million
positive user feedback increase (graph to Claude 3 Haiku)13%
Positive feedback baseline (graph algorithm)58%
positive feedback with Claude 3 Haiku71%
Show all 9 reported metrics
personalised titles generated in 2025over 5.5 million
positive user feedback increase (graph to Claude 3 Haiku)13%
positive feedback baseline (graph algorithm)58%
positive feedback with Claude 3 Haiku71%
positive feedback with Amazon Nova Pro73%
negative feedback with Amazon Nova Pro12%
time to first suggestion reductionfrom 1.41 seconds to 0.92 seconds
time to first suggestion speed improvement35%
engagement and purchase rate improvementmeasurable uplifts in engagement and purchase rates
Reported stack
Amazon BedrockAmazon Nova LiteAmazon Nova ProClaude 3 HaikuConverseStream APIFastAPIconvolutional neural networks
Source
https://aws.amazon.com/blogs/machine-learning/how-popsa-used-amazon-nova-to-inspire-customers-with-personalised-title-suggestions/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Switching to retrieval-augmented generative AI improved positive user feedback by 13% over the graph algorithm with Claude 3 Haiku, and Nova Pro further lifted satisfaction to 73% positive feedback.

What tools did this team use?

Amazon Bedrock, Amazon Nova Lite, Amazon Nova Pro, Claude 3 Haiku, ConverseStream API, FastAPI, convolutional neural networks.

What results were reported?

Personalised titles generated in 2025: over 5.5 million; positive user feedback increase (graph to Claude 3 Haiku): 13%; Positive feedback baseline (graph algorithm): 58%; positive feedback with Claude 3 Haiku: 71% (source-reported, not independently verified).

How is this ecommerce ops AI workflow structured?

User selects photos, app extracts metadata → Geocoding and subject classification → Retrieve few-shot examples from database → LLM generates title-subtitle-category triplets → Stream parsing and validation → Customer feedback loop.

Related ecommerce ops cases