FLITTO

Flitto is a Seoul-based AI translation company building large-scale multilingual training data for machine translation models. From late 2022 through 2023 I was their Translation Project Manager for English-to-Arabic, the language pair their model was weakest in and the one I was strongest in. The job was not translation itself. It was running the team that produced the data the model would learn from.

Day to day, I managed a small team of Arabic translators, set scope and acceptance criteria on incoming projects, reviewed and corrected output, and coordinated directly with Flitto's product owner on what the model needed next. The work taught me what a translation quality bar looks like when the consumer is not a human reader but a model that will learn the bar as a pattern. A single inconsistent rendering, repeated across enough examples, becomes a learned mistake. The job is to not feed mistakes to the model.

It was also the first time I worked on the production side of AI rather than as a downstream consumer of it. The questions you ask change when the work goes into training data instead of into a human's reading experience. Not "does this sentence land?" but "will the model generalize correctly from this sentence?" Not "is the register right?" but "is the register marked clearly enough that the model can learn it?" That shift in vantage was the most valuable part of the year.

Two years later when I started building The Architect AI, that view from inside the data pipeline was already part of how I thought about the problem. Training quality bounds output quality. The team running the data is part of what makes a model work or not. That is engineering.