A user spends hours meticulously selecting the best memories from a summer trip to Provence, arranging layouts and adjusting filters, only to finish the project with a title like France 2024. This is a recurring friction point in the digital photobook industry. Most customers are not professional copywriters, and the resulting generic titles often clash with the high-end aesthetic of the physical product, undermining the overall design quality. The gap between a curated visual experience and a bland textual label creates a subtle but persistent drop in the perceived value of the final product.
The Architecture of AI-Driven Personalization
To solve this creative bottleneck, Popsa redesigned its title recommendation engine by leveraging Amazon Bedrock and the Amazon Nova family of large language models. The goal was to move beyond generic suggestions and create a system capable of generating brand-aligned titles and subtitles in 12 different languages. By utilizing the unified API of Amazon Bedrock, the development team was able to conduct rapid A/B testing, swapping between Anthropic's Claude 3 Haiku and various Amazon Nova models, including Nova Lite and Nova Pro, to find the optimal balance of creativity and performance. This infrastructure allowed Popsa to generate over 5.5 million personalized titles throughout 2025.
Ensuring the quality of these millions of outputs required a rigorous evaluation framework. The team built a dedicated pipeline using a dataset of more than 100 photobook examples, employing an LLM-as-a-judge methodology. In this setup, a high-capability model acts as the evaluator, scoring the outputs of other models based on three critical dimensions: adherence to brand guidelines, grammatical precision, and contextual accuracy relative to the user's photos. This automated judging system eliminated the need for manual review of every iteration, allowing the team to iterate on prompts in real-time.
Central to the system's accuracy is the implementation of retrieval-based few-shot prompting. Rather than relying on a static prompt, the system queries a database for successful title examples that match the specific design style of the current photobook. These examples are fed into the model using a structured user and assistant message format, effectively teaching the model the desired pattern before it processes the actual user data. While Claude 3 Haiku provided a strong baseline, the Amazon Nova suite offered a significant advantage in global scalability, supporting over 200 languages while maintaining lower latency for the end user.
From Rigid Graphs to Fluid Intelligence
For years, the industry standard for this feature was the rule-based algorithm, which Popsa referred to as the Title Suggestion Graph. This system operated on a set of rigid templates. If the metadata indicated that all photos were taken on a single date, the graph would trigger a predefined suggestion such as On this Day, followed by the specific date as a subtitle. While functional, this approach was fundamentally limited by its lack of creativity; it could only recognize patterns, not interpret meaning. The results were predictable and often felt robotic, failing to capture the emotional essence of the user's memories.
The transition to generative AI fundamentally altered the data pipeline from extraction to execution. When a title recommendation is requested, the system first decrypts the timestamps from the design file. It then performs reverse geocoding to convert raw latitude and longitude coordinates into recognizable addresses and city names. Simultaneously, a Convolutional Neural Network (CNN) analyzes the images to classify landmarks and objects. This multi-modal data is then synthesized into a descriptive prompt for the LLM. Instead of a date, the model receives a rich context: a ski photobook containing 21 photos taken between January 21 and 23, 2025, in the Alps.
This shift from template-based logic to contextual intelligence produced measurable business results. After transitioning from the graph algorithm to Claude 3 Haiku, the rate of positive user feedback climbed from 58 percent to 71 percent, a 13 percentage point increase. Multivariate testing across hundreds of thousands of users further revealed that AI-generated titles significantly improved both the design completion rate and the final purchase conversion rate. Currently, Popsa is further optimizing this pipeline by using Amazon Bedrock to switch model IDs, allowing them to validate Nova's performance against cost and speed metrics without rewriting the underlying integration code.
The success of this implementation demonstrates that the real value of generative AI in e-commerce lies not in simple text generation, but in the orchestration of data extraction, retrieval-augmented generation, and brand-specific constraints to drive actual commercial conversion.
