I'm a localization program manager and engineer. I have an MA in Translation and Localization Management from MIIS ('26), and I've worked as a localization PM at TikTok and a localization engineer at Lionbridge. I build AI and i18n tooling that makes localization faster and catches problems earlier.
These two projects are built on public Salesforce material and are designed to work together.
| Project | What it does | Result |
|---|---|---|
| llm-localization-pipeline | An EN→ZH LLM pipeline. It feeds the model glossary and translation-memory context, checks every output with a code-based QA gate, and hands flagged segments to human review as XLIFF. I benchmarked 6 configurations. | With context, chrF rose by 10.3 points, and UI labels matching Salesforce's established translations went from 77% to 100%, at $0.12 per 1K words. AI edit and verify stages tripled the cost for only +0.3 chrF, so the recommended setup leaves them out. |
| lwc-i18n-readiness | An AST-based i18n linter and GitHub Actions PR gate for Salesforce Lightning Web Components. It also does pseudo-localization and a Custom Labels ↔ XLIFF round trip. | On held-out Salesforce sample apps it caught 125 of 133 hardcoded strings, with zero false positives. |
How they connect. I ran a preregistered, blind-judged experiment that sent real Salesforce UI strings from the linter's label tooling through the AI pipeline. Single strings translated well with or without context. Every failure came from a sentence assembled from fragments or from a hardcoded unit, and a per-string QA gate can't see either problem. These fixes belong upstream, before translation starts. Read the experiment →
- Comment-Replacer (live): a browser-only tool that applies reviewer comments in .docx files. At CLEAR Global it cut about 85% of the manual effort on a recurring linguist task.
- AI-short-drama-maker (live): a web app that runs LLM, image and video generation APIs through a 6-stage production pipeline.
- Vendor-Performance-Dashboard (live): a prototype that scores translators, PMs, LQA reviewers and MT specialists on metrics that fit each role, and flags patterns such as MT rubber-stamping or on-time delivery that hides rework. It uses simulated data.
I build with Claude Code as an AI pair programmer. The scope, the design decisions, the linguistic review and the evaluations are my own.