I lead Material Bank’s AI/machine learning team. We create agents for designers, brands, finance, legal, customer service, sales reps, etc. Please reach out if you are interested in building LLM products, Agentic systems, recommendation systems, 3D rendering systems, etc.

Different from my previous customer-facing e-commerce-tech, I now use machine learning, reinforcement learning to solve problems in time series forecasting and combinatorial optimization in fulfillment side of retail, to grow my knowledge and complete my e-commerce journey.

With deep learning, large language models, multi-modal encoders and reinforcement learning, as a technical leader and impact influencer, I help saks.com transform its onsite experience of Search, Recommendations and Personalization to improve online customer satisfaction measured by conversion, overall GMV and profit. Conduct research on new methods that are unique to fashion and luxury retail domain. With the team, I achieved 9% add-to-cart rate increase (equivalent to conversion/revenue) within the first 2 months with our revenue driven product ranking algorithm that considers both product relevance and business impact; leading the team to identify significant opportunity in product search and personalization, and drive technical roadmaps and being hands on with the team for the implementation of the improvements. Coming from the research world, my background enables me to think critically and out of the box, while being realistic and mindful about the timelines and trade offs to achieve business goals. I also build strong collaborations with other technical teams. Growth mindset, always improving, relentlessly pursuing better solutions to improve business goals and team efficiency. Laser focused on business outcome and in the meantime being an encouragement to the team and improve work efficiency by driving us towards the goal.

Driving NLP and ML capabilities for Wayfair Search and Recommendations. My idea on the implementation of the loss function for embedding based search led to $670M expected yearly revenue gain. Using my knowledge of language encoders has also identified opportunity for query reformulation, autocomplete and query understanding.

Using reinforcement learning and neural language model, I created a bot that plays word games with human. Bot is trained with self play algorithm with human knowledge as regularization. The bot was brought on line as IBM guess-the-word game, and the data from human feedback is used to improve the agent quality, a regime similar to instructgpt/Chatgpt training. Our bot and our study about human behavior won CHI 2020 best paper award. Studies about language models also got published on AAAI to confirm language models’s ability to encode common sense knowledge. Before that I was one of the early teams that joined SQUAD dataset ‘competition’ and apply retrieve-and-read approach for open domain question answering. Paper got published on AAAI ICLR and arxiv.

Team lead on Neural machine comprehension and question answering. We started to work on neural models for open domain QA. Our study’s on large n-gram language models on problems such as unsupervised spellchecking also proved larger dataset can get ‘free lunch’ for model performance. Early explorers of memory-based neural networks such neural Turing machines and memory networks.

Research assistant on machine learning and NLP. Created geolocator, a tool that identifies locations in social media text, under funding support.

Graphical Models for tackling retail bank marketing problem.

Information RetrievalInformation extractionRecommender systemsOntology constructionUser modeling

Word Sense Disambiguation, won SemEval 2007 task 11. People’s scholarship.