Curriculum vitae

Experience

May 2025 to present

Co-founder and CTO, FluenTea Berlin and Cincinnati (US C-Corp). Video classroom and AI speaking coach for language schools.

• Co-founded FluenTea, a video classroom with an AI speaking coach for language schools, and built the product alone as the technical co-founder, from the first commit to Google Play. Bought managed real-time video, a hosted speech model and a managed database so that one person could take it from concept to launch in three months and keep a practice call at about 1.4 cents a minute.
• Designed the voice agent's tool use with safety first and tested it the way a research system is tested. Every tool is listed in a manifest that tests enforce, each role sees its own list, nothing that reaches another person happens without a confirmed second turn, and text the learner did not type is treated as data. A red-team suite, a second model grading real answers against a baseline per error type that each release has to match or beat, and weekly reads of real transcripts.
• Ran production alone on Kubernetes for eighteen months, with deploys behind a test gate and a live-schema check, CI on seeded fixtures, contract tests for route authentication and security headers, and property-based timezone tests added after a school entered its timetable a day late.
• Worked directly with teachers, coordinators and a non-technical team. Rebuilt the dashboard from user interviews (monthly active users up 247% between April and August 2026), then had ChaCha run the interviews herself: eight scripted questions by voice or text, the server deciding when an interview is complete, eleven fixes from the first round. Built the age gate, parental consent, GDPR deletion, recordings, attendance and per-student speaking time a university required before enrolling minors.
• Built the tooling around the product: a nightly scripted walk-through that signs up and uses the app as a confused first-time user, Microsoft Clarity for rage-clicks and drop-offs, weekly reads of ChaCha transcripts for complaints, AI-drafted school outreach by email and LinkedIn with a do-not-contact list, and a written plan-build-test workflow with coding agents that turned eleven interview items into code in a week.
• Launched Sep 2025, three months after the concept; more than 16,000 registered users from 136 countries; three university clients, the first a university in Mexico with 500 students; an Android app on Google Play. Consumers signed up in large numbers and schools were the ones who paid, so the 2026 roadmap followed the schools.

Apr 2021 to Nov 2026

Applied Scientist, computer vision Amazon, Berlin

• Technical lead on three production launches on amazon.com and Alexa with multi-million dollar impact from increased sales and reduced returns; five papers at Amazon's internal computer vision conference (one oral, one best-paper shortlist) and D-Extract at WACV 2023.
• Dimension extraction from product images: conceived a text-VQA approach (a TAP-style multimodal transformer that reads measurements, arrows and tables in the image), wrote the annotation protocol and ran a five-person annotation team across 130 product types, productionised the model as an inference pipeline that backfilled the catalogue, and led the public dataset release on amazon-science. Sizes filled in for more than ten million products; Door Desk award.
• Interest and theme prediction for social shopping images and video: zero-shot and low-shot CLIP (open_clip ViT-L) against a Swin baseline with a custom loss for missing labels, Tip-adapter and class-description augmentation, four times faster inference than Swin. Designed the crowd annotation (qualification tasks, model-proposed candidates) over 270 interests and 30,000 videos and the audit-based per-interest thresholds at 90% precision. Launched into Amazon's interest mapping and used by two recommenders.
• Visual similarity and refinement for Alexa Shop the Look: embedding retrieval with sibling and near-duplicate filters for diverse results, index coverage raised from 15% to 95% by tracing the data pipeline to one attribute, and a CLIP-style composed-retrieval model (reference image plus modification text). Launched at 100% in June 2023; shown at re:MARS.
• 2025: generative video and 3D feasibility studies: frame selection for structure-from-motion and Gaussian splatting, video transitions and product swap with diffusion models, image editing with Flux Kontext.
• Supervised four research interns; organised the Berlin ML Workshop 2023 (180 attendees); won the Berlin GenAI hackathon 2024 with auto-generated buying-guide videos; Amazon Berlin Champion of Diversity and Inclusion award, 2021.

Mar 2019 to Mar 2021

Research Staff Member IBM Research Almaden, San Jose, California. Scalable Knowledge Intelligence department.

• Deep learning and computer vision for enterprise document extraction and understanding: algorithms to find and read tables, and active-learning tools that let users build models for their own documents.
• Technical lead and main developer of GTE, an end-to-end table extraction system (WACV 2021), and of TableLab, the labelling interface built on it, from concept to launch.
• Lead organiser of SemEval 2021 Task 9, SEM-TAB-FACTS, on fact verification against tables. Co-author of the CORD-19 dataset (ACL 2020 workshop) and of D2S, document-to-slide generation (NAACL 2021). Tutorials on table extraction at ICDM 2019 and VLDB 2020.

Sep 2014 to Dec 2018

Graduate Research Assistant University of Washington, Seattle. GRIDLab, Neural Systems Lab, Brunton Lab. Advised by Bingni Brunton and Rajesh Rao, with Jeff Ojemann and Ali Farhadi.

• Sensor-fusing deep learning models that decode and predict human behaviour from hundreds of hours of simultaneous video, audio and electrocorticography recordings (AAAI 2018, oral).
• Founded the project to automate the annotation of naturalistic neuroscience data; it became a multi-year effort funded by NSF and DARPA, and parts of the dataset were released.

Jun to Sep 2018

Computer Vision PhD InternZillow, Seattle. Computer vision for housing products.

Mar to Jun 2016

Machine Learning PhD InternRakuten Institute of Technology, Tokyo. Unsupervised user profiling for search ranking and page design.

Education

Dec 2018

PhD, Computer Science and EngineeringUniversity of Washington, Seattle

May 2014

BSc (Honours), Computer ScienceUniversity of British Columbia, Vancouver. Dean's Honour List every semester.

Selected publications

• Ghosh, P., Wang, N., Yenigalla, P. D-Extract: Extracting dimensional attributes from product images. WACV 2023.
• Sun, E., Hou, Y., Wang, D., Zhang, Y., Wang, N. X. R. D2S: Document-to-slide generation via query-based text summarization. NAACL 2021.
• Zheng, X., Burdick, D., Popa, L., Zhong, X., Wang, N. X. R. Global Table Extractor (GTE): A framework for joint table identification and cell structure recognition using visual context. WACV 2021.
• Wang, N. X. R., Mahajan, D., Danilevsky, M., Rosenthal, S. SemEval-2021 Task 9: Fact verification and evidence finding for tabular data in scientific documents (SEM-TAB-FACTS). SemEval 2021.
• Wang, L. L. et al. (incl. Wang, N. X. R.) CORD-19: The COVID-19 open research dataset. ACL 2020 Workshop on NLP for COVID-19.
• Wang, N. X. R., Farhadi, A., Rao, R., Brunton, B. AJILE movement prediction: Multimodal deep learning for natural human neural recordings and video. AAAI 2018, oral, 25% acceptance.
• Wang, N. X. R., Olson, J. D., Ojemann, J. G., Rao, R. P. N., Brunton, B. W. Unsupervised decoding of long-term, naturalistic human neural recordings with automated video and audio annotations. Frontiers in Human Neuroscience, 2016.

Full list on Google Scholar

Last updated November 2026.