01medium.com
From Activity to Intent: Generating User Journeys with LLMs
Pinterest replaced its multi-stage journey-clustering system with a fine-tuned 4B Qwen3 model that converts up to 360 days of chronological activity into ranked, locale-specific intent names. It uses frontier-model synthetic data, supervised fine-tuning, and LoRA; the 4B model balances quality and throughput. Production serving with NVIDIA Dynamo, vLLM, and L40S GPUs sustains about 775 requests per second with 1.2-second median latency. Online experiments improved email click-through by 1.1% and push opens by 1.3% versus clustering.