The system we inherited was a batch pipeline running on scheduled cron jobs. Every night, 2.1 million positions were recalculated from scratch. The AI models themselves were solid — the infrastructure around them was what was failing.

We introduced a streaming architecture using Apache Kafka for position updates, with inference served through a FastAPI layer backed by Redis. The OpenAI API calls were replaced with a fine-tuned model running on-premise — lower latency, lower cost, and no data residency concerns.

The result was deployed to staging in week four and to production in week seven. It has run without a production incident since launch.