Working

Here is a list of past working experience.

Lira
AI Tech Lead
Dec. 2022 - Jul. 2023
- Data Science Part-Lead the lip reading project
- Design several models for the lip-reading task, including Transformer-backbone-based, ResNet-backbone-based, and several pure-video based.
- Build the model, in the latest work, and subtraction based transformer-backbone model achieved the state of the art performance in the lip reading task.
- Lead the deployment of model transfer to app platform(in process).

- Data Engineering Part-Lead the database building and data annotation
- Managing the Azure collected lip reading dataset
- Build and manage the Azure storage for the collected data

Merck
Data Science Intern
Jan. 2022 - Jul. 2022
- Built target liability assessment text analyzing model (includes a search engine and a text classifier) based on the paper of target (compound) searching results from PubMed Capabilities of the model: ▪ Extracting all the sentences related to the compound and customizable disease or symptoms

- Built front-to-end prototype of the deep-learning based compound analyzing model (DCM) for Merck historical compound pdf-format-based reports Capabilities of the model: ▪ Parse all the pdf and word files, extracting different sections from those files including the abstract, conclusion, and result.

BaiRong Financial Information Service Company
Machine Learning Intern
Jun. 2018 - Sep. 2018
- Created model for risk control through Logistic regression and Stepwise regression via R and Python; “bad customer” ratio decreased by 5.67%, and payment received ratio increased by 20.3%

- Stacked multiple XGboost into single model, compared with LightGBM; increased AUC from 0.68 to 0.76

Flagship Pioneering FL100 lab
Machine Learning Scientist
Jan. 2024 - Now
- Machine Learning part Leading issue: Alleviating the LLM hallucination through social media validation
- Co-Built a multi-agent system (based on OpenAI) for generating CPG product, the agent system using Async, Instructor and Pydantic which capable to generate a large number any type of structured concept including text and image.
- Built a Vector search solution on AWS for Amazon review database (historical product + review data till 2023), could perform free text vector search for 200-1000 products (scalable) and filtered reviews within 3s.
- Built a zero-shot free text classifier for 1700 nested product categories (4-8 layers each category), based on GPT 4o (not so proud), F1 score achieved 83%, prediction time within 1s
- Built other social media pipelines including Tiktok and Reddit (both requires data authorization), Tiktok was used for predicting social trend using video transcript and reviews, Reddit was using for painpoint generation (like Ideaape: https://ideaape.com/)
- Software developing part
- Majorly working on the backend, leading one workflow which is the social listening workflow and cooperating with the front end (the software is available on Shopify)
- Deployed backend service through FastAPI