Working for LLM safety, Bard/Gemini.

Especially Gemini safety/fairness and Multimodality(Veo and Imagen) safety involving EVAL, and post training (SFT, RL).

Machine learning engr working on Bard and Gemini safety, especially LLM safety and RLHF

*Wearable Life Project for Young Adults with IIDs

• Designed, developed, evaluated a wearable and mobile application for young adults with intellectual and developmental disabilities(IIDs) to help with the coordination and communication between the students and their in-class assistants in inclusive education in a less obtrusive and stigmatizing way, which provided timely intervention notifications in class context sent automatically/manually by assistant to student, including the features of interventions and self-assessment, guided by User-Centered Design, with different relevant stakeholders (students with IID, their assistants, special education experts), including 14 user studies (interviews, focus groups, surveys, usability tests, field studies).

Collected sensor(accelerometer, gyroscope, heart-rate) data form the smartwatch, to recognize the student’s in-class behavior(future work), and I already tried to use deep learning(LSTM+CNN) to do human activity recognition on wearable sensor data(accelerometer and gyroscope).

Developed In Android, SQL. Deep Learning in PyTorch and Python.

Published in CHI’18, ASSETS'21, InterAct'19, ASSETS’ 17.

*Wearable on-task self-monitor for the students with ASD, Jun 2017- Aug 2017

• Designed, developed a wearable app to let the students with ASD self-monitor their on-task daily routine on their smartwatch (check on-/off-task). There are 4 kinds of self-monitor tasks (in-class, off-class, weekday, weekend), which their support staff or their parents can program the content and repeat pattern of the task on the phone. In Android.

*Log Parse for a cognitive intelligence analysis assistant Jun 2018 – Aug 2018

• Designed, developed a statistical analysis program for Cogent (cognitive assistant for the intelligence analyst) project. Parse XML logs of Cogent. Analyze hands-on and usage of various operation(total time, frequency, what operation used and never used by a user per problem), usage and trend line of Help operations. Funded by IARPA. In Java.

*Mental Stress Level Recognition through the Wearables, Ap 2017- May 2017

• Analyze the data collected from Amulet watch. Extract features from the heart rate sensor and accelerometer. Learn an SVM model to predict the stress level. Funded by NSF. In MATLAB.