Stackline · OpenAI Consulting Services

“Our GPT integration went from a risky experiment to the most reliable part of the product. The guardrails and evaluation work made all the difference.”

James Whitfield, Chief Technology Officer at Stackline

James Whitfield

Chief Technology Officer, Stackline

London, United Kingdom

99.9%

AI feature uptime

<0.5%

Hallucination rate in production

7wk

From prototype to production

The situation

Stackline's prototype AI assistant was the hit of every sales demo — and the terror of every engineering review. It hallucinated confidently, broke under edge cases, and James couldn't justify shipping something he couldn't measure, let alone guarantee.

The board saw the demo and wanted it live. James saw the logs and wanted it safe. Both were right.

The partnership

JSXDOM rebuilt the feature the way production AI has to be built: a RAG pipeline grounded in Stackline's own documentation, a guardrail layer constraining what the model could claim, and — the piece nobody demos — an evaluation harness that scores every release against hundreds of adversarial test cases.

Nothing shipped until the evals said it was ready. Then it shipped, and the evals kept watching in production.

The outcome

The assistant is now Stackline's most-used feature and its most reliable one. Hallucinations run below half a percent, uptime matches the core platform, and James ships model updates with the same confidence as any other code — because the evaluation suite won't let him do otherwise.

Project: OpenAI Consulting Services — delivered by JSXDOM for Stackline.

Explore this service

Your story could be the one we tell next. Ready?