Machine learning · Computer vision · LLM systems
I build real-time voice agents and the test suites that measure what they say. Before that I built computer vision models that run on amazon.com.
PhD from the University of Washington, then IBM Research, five years at Amazon, and since 2025 co-founder and CTO of FluenTea, where I designed and built the product, from the voice agent to the Android app.
FluenTea is a video classroom and AI speaking coach for language schools. I co-founded it in 2025 and built it myself, from the first commit to Google Play. Every class and practice call is transcribed. From the transcripts we compute how long each student spoke, which grammar they used and where they struggled, and what the teacher should go over again next lesson.
The coach, ChaCha, follows a written behaviour spec that decides how hard she pushes a learner depending on their level and how they are doing. Before she went into classrooms I built tests for her at three levels. The first set checks what she was told before she answered. The second feeds her text written to make her misbehave. In the third, a second model grades her real answers against a baseline per error type, and a release that scores worse does not ship. In the first university pilot, most students kept practising with her for longer than their teachers had asked. She also runs our user interviews now, calling students and teachers and asking eight scripted questions, and the answers go into the feedback tracker.
The rest of the product is the web app, the backend and the Kubernetes cluster, the Android app, payments, and the consent and safety features a university required before enrolling minors. We launched three months after the concept. There are now more than 16,000 registered users in 136 countries and three university clients, the first of them in Mexico with 500 students.
Before that I was an applied scientist at Amazon in Berlin for five years, working on computer vision and multimodal models that run on the retail site, and a research staff member at IBM Research Almaden, where I worked on getting machines to read tables in documents. My PhD was about decoding what people are doing from recordings of their brains and video of their behaviour.
In the first university pilot, most students kept practising with the coach for longer than their teachers had asked. I co-founded FluenTea and built the whole product myself, from the voice agent to the Android app. Language schools run live classes in it, the coach runs practice calls between classes, and teachers are told what to reteach.
Technical lead on three computer vision projects that went to production on amazon.com and Alexa. One reads product dimensions out of images with a text-VQA model and filled in sizes for more than ten million products; another predicts the interests in creators’ photos and videos with CLIP and feeds two recommenders. Door Desk award; D-Extract at WACV 2023.
At IBM Research Almaden I led GTE, a table extraction system (WACV 2021), built the TableLab labelling tool around it, and organised a SemEval 2021 shared task on table fact checking. My PhD work at the University of Washington, decoding behaviour from brain recordings and video, was an oral at AAAI 2018.
I love brains and computers, basically any cognitive system. I was born in Nanjing, moved to Canada at nine, studied in Vancouver and Seattle, worked in San Jose and Tokyo, and have lived in Berlin since 2021, where I am slowly learning German. I like living inside other cultures, which is probably also why I like role-playing and strategy games (Dragon Age and XCOM most of all).