We rewrote JSONata with AI in a day, saved $500k/year

This post describes how Reco rewrote their JSONata JavaScript evaluator in Go using AI-assisted development, eliminating expensive RPC overhead and achieving $500K/year in cost savings.

• •The original setup used jsonata-js pods on Kubernetes, requiring Go services to call Node.js processes over RPC at ~150 microseconds per call, costing ~$300K/year in compute
• •Using AI to port the official JSONata test suite to Go, a 13,000-line pure-Go implementation (gnata) was built in 7 hours for $400 in tokens with 1,778 passing test cases
• •gnata uses a two-tier evaluation architecture: a fast path for simple expressions evaluated directly on raw JSON bytes with zero heap allocations, and a full path with a complete parser for complex JSONata 2.x expressions
• •A streaming layer (StreamEvaluator) merges all field paths from N expressions into a single scan and caches evaluation plans immutably per event schema with a lock-free hot path