We fine-tuned an AI model for log search – Accuracy 50% to 80%

A team fine-tuned an AI model specifically for log search after discovering that general-purpose models fail catastrophically on operational data, confusing critical failures with successes and missing causal relationships between incidents. Through contrastive learning on 78,000 training pairs from real enterprise logs and incident scenarios, they improved search accuracy from 50% to 85% and achieved 95% correlation with human expert rankings, enabling engineers to actually find root causes during outages.