Skip to content
Stephanie De Luna
Lead AI Product Designer · Meta Wearables
Back to work
Case Study

i18n Internationalization

From Spanish localization to the guidelines that shaped wearable AI across languages

Authored the localization guidelines for voice interactions on Meta's wearable devices, starting with Spanish. These guidelines scaled across the ecosystem, shaped ML models, and led to me joining the Translations team.

Vision

A global assistant that understands everyone and fosters equitable participation in the future of technology.

Starting with Spanish

Native expertise as a design advantage

Localization isn't translation. As a native Spanish speaker, I defined what "natural" sounds like in a non-English voice experience, adapting interaction patterns, tone, and cultural context across Meta Portal and Ray-Ban Meta.

Beyond translation

Adapted phrasing structures, formality levels, and conversational rhythms, not just words.

Cultural context

The assistant had to feel like it belongs in the user's language, not like a translated English experience.

Scaling the Guidelines

Training translators on conversation design

English designs went through an automatic pipeline to third-party translators who had no conversation design context. Translations lost their conversational quality, making voice interactions feel unnatural in other languages. I trained translators on CxD guidelines, significantly improving quality and supporting English feature parity across global markets.

v1
Literal translations

No conversation design context. Translated strings felt stiff and unnatural.

Trained translators on CxD
Pushed for
Design-informed translations

Translators trained on CxD guidelines. Quality improved significantly, supporting feature parity.

Fixing the Pipeline

Improving the review process

The review process was slow, requiring multiple steps and links to request a single change. I streamlined the pipeline so designers could review and request changes faster.

TTS listening tool

Used an internal tool to hear TTS output on Ray-Ban Meta, critical for evaluating if translations sounded natural when spoken aloud.

Streamlined review

Simplified the multi-step process so designers could review and request changes in fewer steps.

From Guidelines to Models

Shaping how models learn language

As the team shifted to model-based systems, these guidelines became essential for training and evaluating ML models. Collaborated closely with linguists, PMs, and ML engineers to ensure models reflected the same quality as hand-crafted flows.

Linguists

Co-drove localization initiatives, combining design expertise with linguistic precision.

ML Engineers

Guidelines became training data and evaluation criteria for model-based voice interactions.

PMs

Aligned localization quality with product goals and language expansion timelines.

Outcome

Guidelines scaled from Spanish to every language across Meta's wearable ecosystem, shaping both human-designed flows and ML models. This i18n expertise led to joining the Translations team, where I applied those learnings to design live translation on Ray-Ban Meta.