Ramp's iOS team built an automatic receipt-matching feature that scans a user's photo library and links receipts to transactions entirely on-device. Starting with basic string comparison (too brittle), they moved to Apple Intelligence's FoundationModels API for a single LLM call (too prone to hallucination), then iterated to a hybrid approach: using deterministic iOS data-detector APIs for amounts and dates while relying on the LLM only for fuzzy merchant-name matching. Rigorous evals guided each iteration. The final v3 achieved 87% precision and more than doubled recall, and is now shipping to all Ramp users.
Table of contents
Version 1.0: the moonshotReality checkWWDC 25Version 2.0: Apple Intelligence to the rescue!EvalsVersion 2.1: Apple Intelligence, refinedReality check, for realVersion 3.0: the best of all worldsConclusion2 Impressions