Israeli Startup Saves $500k Annually Using AI-Built JSONata Engine

Reco, an Israeli SaaS security firm, has rebuilt its JSONata expression engine using AI, reducing annual infrastructure costs by approximately $500,000 and improving processing speeds across its data pipeline. 

The new system, called gnata, replaces a JavaScript-based implementation that previously ran as a separate service.

The original setup required Go services to call Node.js processes over remote procedure calls (RPC) to evaluate JSONata expressions, a query language for JSON data. 

This architecture introduced latency and scaling challenges. 

Each evaluation incurred a network round-trip of around 150 microseconds before execution began, adding overhead at scale.

Reco reported that the system processed billions of events across thousands of expressions, with infrastructure costs reaching about $300,000 annually. 

“For a simple field lookup like user.email = ‘[email protected]’—something that should take nanoseconds—we’re paying microseconds just for crossing a language boundary,” wrote Nir Barak, Principal Data Engineer at Reco.

The company adopted a method similar to an earlier Cloudflare experiment called ‘vinext’, using AI to implement software against an existing specification and test suite.

Engineers ported the official JSONata test suite to Go and iteratively generated code until all tests passed. The process took seven hours and cost $400 in compute tokens, producing a 13,000-line implementation with 1,778 passing test cases.

The resulting system runs as a native Go library, eliminating the need for inter-service communication. 

It introduces a two-tier evaluation model, where simple expressions are processed directly on raw JSON bytes without parsing, while complex queries use a full parser. 

Reco reported speed improvements of up to 1,000 times for simple operations and 25-90 times for more complex ones.

The rollout included a shadow-testing phase in which both systems ran in parallel. After three days of identical outputs across production workloads, gnata replaced the previous implementation. 

“By the time we promoted gnata, it had already been processing billions of events and producing identical results to jsonata-js,” Barak said.

Further optimisations to the rule engine, enabled by the new architecture, reduced annual costs by an additional $200,000. 

These included batching evaluations and reducing reliance on high-concurrency patterns that previously increased memory and CPU usage.

Barak noted that the project reflects broader changes in software development workflows. “I believe gnata is just the beginning. I suspect 2026 will be the year of surgical refactors,” he wrote.