Automated Multi-Ledger Reconciliation via LLM-Augmented Transaction Classification

The deployment centered on a fine-tuned GPT-4o classification model trained on 14 months of historical transaction data (2.1M labelled exceptions) to predict resolution pathways before human review. A secondary Claude 3.5 Sonnet layer handled unstructured exception notes and cross-referenced counterparty correspondence to surface root-cause patterns. The pipeline was orchestrated via LangChain LCEL with a Pinecone vector index (1536-dim, text-embedding-3-large) storing counterparty behavioural profiles across 6,200 entities. A Redis caching layer reduced redundant embedding calls by 71%. The stack ran on Kubernetes (prod-payments-us-east-1, 8 pods) with LangSmith monitoring classification confidence scores. P99 inference latency settled at 340ms per exception event. The system was deployed in a canary configuration at 10% traffic for 12 days before full rollout at day 34.