Developed backend services in TypeScript for IBM Watson Assistant /
Orchestrate, enhancing a conversational AI platform for seamless
chatbot deployment across channels like Slack, Facebook Messenger, and SMS. Led integration
with MS Teams using Azure Bot Framework.
Migrated microservices from Java (Spring) and JavaScript
to TypeScript to improve maintainability, and transitioned RESTful APIs to
gRPC to boost performance of complex chatbot interactions.
Integrated large language models, such as IBM Granite,
Llama, and GPT-4, into the next generation of the AI agent
product, supporting features like agentic chat mode and function calls, with capabilities
for on-premises LLM deployment.
Customized the open-source Portkey AI Gateway to ensure support for
IBM's proprietary models and cloud environments, enabling applications like
real-time voice conversations (phone channel integration).
Worked with 3 other interns to independently develop a solution for IBM Watson Assistant
to increase the user engagement rate of the product.
Built a new feature with Flask and ReactJS that allows Assistant to generate
interactive help pages
based on the existing conversation flows, which can be embedded into the company's website.
In the user tests conducted with internal and two external clients, the new feature has
increased the user engagement rate by 2.5 times. Also reduced the man power
needed to maintain both the help articles and the chatbot content.
Analyze the huge number of incidents that emerged from
the Azure systems and their mitigation steps, develop
internal tools to automate
the procedures in the life cycle of each incident, detect or even predict
the outage, thereby
improving the stability of the Azure infrastructures.
Responsible for the maintenance and development of an internal tool for
estimating and predicting the radius of the impact of outages.
Optimized the performance of Kusto queries to allow it to handle the alert storm
caused by major accidents, and successfully migrate the tool to the Office
365 team's incident management platform.
Also collaborated on the development of the front-end visualization for our
tools with ReactJS, and the design of the data monitoring and processing pipelines
with
Azure Data Factory.
Developed and maintained a specialized online forum and officially recognized
player community for an indie game "Starsector",
with 7k+ unique visitors per week and 60k+
registered users
Deeply customized an existing PHP and JS-based CMS
to meet the modern security and user experience demands, including a
responsive mobile interface for non-PC users and created
service images for Docker/Podman based deployment.
Use Object Storage Services and CDN network
to optimize the
website and attachments access performance
A visualized web app to preserve and share valuable memories with others
with similar experiences powered by NLP and machine learning, responsible
for the front-end UI design and implementation.
Backend implemented in Python with Flask application server;
Google Video Intelligence
APIs and word2vec were used to extract the features of uploaded videos and
build the
connectivity graph
The frontend is implemented as a single-page app (SPA) built with
ReactJS and bootstrap,
graph visualization implemented with cytoscape.js.
A pattern matching based Danmaku (video comments) filtering tool for
Bilibili.com, a
famous online video-sharing community in China
Using regular expressions to match and filter vulgar
words and
unfriendly speeches, such as racial or geographical discrimination contents
and able to merge and streamline repetitive
contents with edit distance and other NLP algorithms