Visa

Recommending merchants to cardholders is a language problem disguised as a payments problem. The winning architecture attaches a scoring head on top of Llama’s hidden states, fine-tuned jointly with the base model, so a cardholder’s profile and transaction history collapse into one embedding, scored against a learned vector for each of 200,000+ merchants via dot product and ranked by the resulting logits. It beats the best existing model by 16% on Recall@10 and 14% on NDCG@10 with 50x less training data, serving personalization to millions of cardholders. Feeding it meant engineering a distributed Spark ETL pipeline on Kubernetes that turns 76+ billion raw transactions into natural-language training sequences, then 400+ ablation studies across distributed training and inference runs on NVIDIA A100 GPUs to land on that architecture.