Joslyn Lim

A machine learning regression model to estimate the price of motor spare parts for the claim assessor, accelerate the claim processing time, and systematize the part price assessment procedure. Collaborating with business users and the AWS team, I started from a user persona, process understanding, and brainstorming session. The team devised a solution to use machine learning with a web app running on AWS architecture to recommend predicted prices for different car makes and models' spare parts. The final model used was ensemble gradient boosted trees (GBT) regressors. This initiative was expected to improve customer satisfaction as it also shortened the claim processing time and reduced inflation impact on spare part costs.

A classification machine learning model that predicts which policies will likely be discontinued in the upcoming renewal cycle. Regional retail general insurance has a very competitive landscape, so it is imperative to monitor a churn rate and keep existing customers whenever required. The development team collaborated with business users to understand essential features that would be beneficial to use in churn prediction. As part of the data scientists team, I used past years' data to support the claims. The team built a random forest (RF) model to help end users retain the fence-sitters, i.e., indifferent customers and likely-to-churn customers, with incentives or targeted promotions.

The project was designed to predict the named entity, document sentiment and entity sentiment, locality, and emotion with higher accuracy. In 2019-2020, Bert was state of the art (SoTA) for NLP. Still, it was not giving satisfying results for ASEAN languages (e.g., Indonesian, ASEAN Chinese, Malay, etc.) as the underlying models were using mostly US corpus. The team aimed to beat the SoTA for ASEAN languages. From data collection, curation, annotation, QA, data cleaning, model training, evaluation, and testing, the team of 10 was doing end-to-end delivery for our client. I was the owner of one of the language pillars while also helping other language pillars with context understanding for regional languages. The team ran multiple ML experiments and found the transformer model worked the best for a few language corpora and a mixture of them, hence applying the large embedding to multiple language models. As the end deliverable, the team containerized the models and Python pipeline using Docker. The models were able to beat SoTA with around a 0.7 F1 score.

A regional outsourcing and shared service experienced high employee turnover in the past year and wished to reduce attrition, especially for the higher performer. A team of five conducted qualitative and quantitative analytics to understand the employee pain points and proposed a solution to the client. I acted as a data scientist to conduct exploratory analytics using employee attendance and leave, appraisal and performance, bonus and pay, and keyword extraction from interview sessions using Python, R, and Power BI. The insights were then matched with attrition data to predict attrition likelihood with a classification algorithm and summarize (clustering and PCA) the attrition factors so that the client could prioritize and take action on high-impact attrition factors.