Hyper-personalized websites at sub-2-second latency: Adobe's real-time AI architecture

Adobe's principal scientist demonstrates agentic sites that personalize website content in under two seconds per page. The system uses Cerebras hardware and Gemma 4 models, achieving 2,300 tokens per second. The key architectural insight is that personalization happens at the block level, not full page generation, respecting brand constraints.

The demo shows real-time persona adaptation based on browsing signals like time on page and navigation patterns. A coffee equipment site adapts hero cards, product recommendations, and even query results to individual users. The audience-of-one concept promises to deliver unique experiences for each visitor.

The critical tradeoff is speed versus accuracy. Different sites require different model evaluations, and Adobe uses Promptfoo for continuous benchmarking. Cheap, small models often suffice for text generation tasks, making latency the bottleneck rather than capability.

The operational challenge is scale: continuous LLM calls and pre-fetching strategies create significant cost implications. Teams must decide where to pre-generate content and where to compute on demand. The technology works now, but governance, cost management, and brand safety remain unsolved for most organizations.