Customized Deep Learning NLP Library• Conducted research and experiments of latest NLP techniques; chose, tailored, and realized techniques relevant to products, including word2vec, CNN, RNN, Attention, transfer learning, multi-task learning, etc.• Developed a Monte Carlo noise eliminator to detect and correct misclassified data points• Wrapped up above modules as a deep learning NLP Python library to facilitate training of NLP models• Built templates and provided guidance for teammates to learn and use the libraryNLP Features for Human-Resources Products• Communicated with product managers and engineer team leaders to create product roadmaps, and wrote proposals about NLP features according to product requirements • Defined goals of the NLP features and metrics evaluating feature performance• Designed model structures, built customized tools, and collaborated with engineers to determine I/O interfaces• Mentored junior team members in data collection, data processing, model training, and result evaluation• Built API of these models using Flask, wrapped them up using Docker, and deployed the Docker images on internal server for product team to consume

Berkeley Lab is a member of the national laboratory system supported by the U.S. Department of Energy.Worked in Joint Center for Artificial Photosynthesis (JCAP), nation's largest research program for solar-fuels generation.Built a high-throughput pipeline with Python to automatically extract material structures from the Materials Project Database, calculated and analyzed material properties, and screened for possible candidates for water splitting. Found a novel material which outperforms all the traditional materials.

Competition ResultsTop 7% in "Porto Seguro’s Safe Driver Prediction", Bronze Medal;Top 8% in "Quora Question Pairs", Bronze Medal, Team Leader;Top 15% in "Bosch Production Line Performance";Top 15% in "Allstate Claims Severity".

Discovered the dominant role of vdW forces in formation of styrene nano wires by developing kinetic Monte Carlo model and implementing them with Fortran. Revealed that different dendritic ZnO nano structures are resulted by the competition between atom deposition and diffusion by applying Diffusion Limited Aggregation model.