Product Director, Localisation & AI Evals

Architect and ship it end to end, from 0 to 1, across all markets and 20+ languages — where models carry the work and humans set the standard and make the final call. Break the cost curve: hold quality flat while coverage scales across languages, content and release velocity, so cost stops growing linearly with the business. Make the system interoperate cleanly with OKX's existing localisation stack and content pipelines, rather than living as an island.

Stand up an autonomous QA agent that evaluates translation and localisation quality at scale — catching regressions before release, monitoring in production, and self-verifying every fix, using LLM-as-a-judge and eval harnesses rather than manual review. Evaluate not just accuracy and fluency but cultural fit, bias and safety — especially on regulated and financial content, where a wrong or non-compliant phrasing is a real risk. Make quality measurable: define and drive a hard quality signal — human-acceptance rate, or the edit effort needed to reach shippable quality (e.g. Translation Time-to-Edit) — so quality is a metric the system optimises against, not a subjective call. And protect the funnel: conversion on key user journeys should never be adversely affected by language or localisation issues.

Drive a sustained reduction in severe and recurring production issues — every escape becomes a new guardrail, and the system gets smarter with every release.