Developer cuts AI scoring costs by rebuilding LLM pipeline for 10,000 daily job listings

A developer building a production job board platform for a client faced unsustainable OpenAI API costs when scaling an LLM scoring pipeline from 100 test listings to over 10,000 per day. The platform ingests job listings from five ATS providers — Greenhouse, Lever, Ashby, Workable, and Recruitee — normalizing them into a standard schema stored in MongoDB. A critical performance issue caused by MongoDB's skip()-based pagination was resolved by switching to cursor-based pagination using the _id field, eliminating CPU spikes at over one million documents. To ensure consistent, structured output from the LLM, the developer replaced freeform JSON prompts with OpenAI function calling, which enforces a fixed schema for scores and match reasons on every response. The rebuilt architecture addressed both reliability and cost challenges that had made the original approach economically unviable at production scale.