• Lead development of interactive translation system allowing users to choose from output words at each decoding step of Huggingface transformer models• Implemented optimized character-level transformer architecture in PyTorch and pretrained it with masked langauge modeling, outperforming state-of-the-art in Japanese topic classification• Rewrote scientific information extraction pipeline using spaCy, increasing micro f1 by more than 20 points

• Lead NLP engineering (5 engineers) across multiple projects: - Apply LLMs to manga translation, allowing terminology andstory aware translation of >30k manga chapters - Explore training methods to improve LLM translation, pruningnoisy OCR data for monolingual and bilingual fine-tuning - Build LLM pipeline for novel translation with LangChain to partially automate translation of 200 novels - Integrate Mantra’s translation technology into multiple multibillion dollar companies - Train multilingual and terminology-aware translation modelswith Fairseq, improving BLEU on multiple languages by >10 - Work with professional translators to improve evaluation• Develop and maintain application backend using Django and AWS• Rewrite flagship mobile application from scratch in Flutter

Research into deep learning approaches to retrieval-augmented and explainable grammatical error correction

• Lead development of, tested, and maintained sole ownership of Scala distributed scheduler that powered proactive messaging for eBay’s chat bot and Chinese iPhone app• Research implementing bootstrapping NLP process to extract informative noun/adjective pairs from product reviews was accepted for presentation at eBay’s internal research conference (25% acceptance rate)• Developed Ruby on Rails web service that allowed non-technical staff to write and schedule proactive messages from chat bot to users

• Rearchitected the pipeline used to import shipping charge data from shipping partners, increasing the accuracy of international shipping prices on eBay’s search results page by up to 40%• Maintained and developed new features in Java/Spring services at scale (several million requests per hour), wrote tests using JUnit and Mockito• Drove an effort to automate excessive manual operations work, saving team tens of hours of work per week

Owned and expanded functionality of Ruby on Rails application automating Hadoop permissions changes and other operational tasks