Popsa generates 5.5M personalized photo book titles using Amazon Nova, cuts costs with 73% user satisfaction

Popsa, a photo book technology company serving 50+ countries in 12 languages, has generated over 5.5 million personalized photo book titles in 2025 using Amazon Bedrock and the Amazon Nova model family.

Migration from rule-based to AI-generated titles

Popsa initially launched its Title Suggestion feature in 2021 using a rule-based algorithm called Title Suggestion Graph. This system analyzed photo metadata (timestamps, geocoordinates) and on-device computer vision features to generate titles following predefined templates.

In June 2024, the company migrated to generative AI using Claude 3 Haiku with retrieval-based few-shot prompting. This approach:

• Retrieved similar photo book examples from a database
• Used examples as conversation context before generating new titles
• Enforced strict constraints: 36-character limit, valid JSON format, predefined category matching

The Claude 3 Haiku implementation increased positive user feedback from 58% to 71% compared to the graph-based method, according to Popsa.

Amazon Nova deployment results

In early 2025, Popsa tested Amazon Nova Micro, Lite, and Pro against Claude 3 Haiku through multivariate A/B testing with direct in-app feedback collection.

Model performance results:

• Nova Pro: 73% positive feedback, 12% negative (highest quality)
• Claude 3 Haiku: 71% positive feedback
• Nova Lite: Near-identical quality to Claude 3 Haiku at lower cost and faster response times
• Nova Micro: Outperformed legacy graph method but lagged other LLMs

Popsa selected Nova Lite for production deployment based on the quality-cost-latency tradeoff. The company reports Nova Lite delivered "near-identical quality to Claude Haiku at lower cost and faster response times," though specific pricing and latency numbers were not disclosed.

Technical architecture

The Title Suggestion Service processes requests through multiple stages:

1. Decrypts and extracts timestamps from user designs
2. Performs reverse geocoding on GPS coordinates
3. Classifies photo subjects using object detection
4. Generates descriptions (e.g., "A skiing photobook with 21 photos taken in the Alps between 21st January 2025 and 23rd January 2025")
5. Applies retrieval-based few-shot prompting to produce title suggestions

The system operates across 12 languages and enforces output constraints including character limits, valid JSON formatting, and category classification for icon rendering.

Evaluation methodology

Popsa built an evaluation pipeline using:

• A dataset of 100+ example photo books
• Automated metrics: character limit compliance, valid category percentage, JSON format validation
• LLM-as-a-judge for brand style, theme consistency, and title-subtitle cohesion
• Multivariate testing with hundreds of thousands of users
• Direct user feedback ratings (positive/neutral/negative)

The company tracked downstream metrics including Design Created and Purchase rates, both of which improved according to Popsa.

What this means

This case study demonstrates practical LLM deployment optimization at scale. Popsa's methodology—automated guardrails, LLM-as-judge evaluation, and multivariate testing with real user feedback—provides a template for product teams evaluating model migrations. The ability to test multiple models through Amazon Bedrock's unified API allowed Popsa to switch from Claude 3 Haiku to Nova Lite in hours rather than weeks, according to the company. The 5.5 million titles generated in 2025 indicates significant production volume, though Popsa did not disclose what percentage of its user base actively uses the feature or comparative costs between Claude 3 Haiku and Nova Lite deployments.