Using LLMs for Workday Skills Cloud.

Worked with a newly-formed team to create an ML recommender system for Workday prompt items, from idea to product launch.- Created an end to end ETL pipeline in Airflow: - Detect that raw user interaction data is ready in S3 bucket. - Run ETL (Scala fat jar) on EMR to turn into format for knowledge graph. - Run Pyspark training pipeline on knowledge graph data. Select best models and output to s3. One model, matrix factorization, creates user/item embeddings - send to vector index in GCP for ANN search. - The Airflow pipeline had Slack notifications, and logs/metrics were sent to Cloudwatch, S3, and Wavefront. - All python code ran in custom python containers with a custom Kubernetes pod operator.- Collaborated on the recommendation serving layer, and designed a strategy for warm/cold users/items and fallback strategies if the vector index was unreachable. Created end-to-end test models and bash scripts to test recommendations in all Workday datacenters.- Worked with devops to deploy (Ansible) the ETL pipeline and serving layer, and created a devops runbook. Worked on-call with Pagerduty. Quickly fixed bugs in early deployment.- Created the code repo. Repo had git hooks, linting, Jenkins CI, slack integration, a python poetry env, initialization scripts for AWS ec2 instances, config using hydra. The CI ran unit tests and sent containers to a docker registry. The repo also had setup scripts for dev environments in Jupyter for python and scala.- Regularly code-reviewed teammates.- Read machine learning papers and discussed with the team. Led data science experiments. Regularly discussed and revised design specs with team, considering customer data security and product performance requirements. Met with UI team to discuss api spec for customer interaction collection, and for recommendation requests.- Conducted interviews solo and with the team: technical code pair, project retro, api design, ML fundamentals, behavioral, team debrief.

Rotation Program:---------------------Rotation 1: Front EndLanguages and Tools: XpressO (Workday proprietary)- Designed a Receivables Aging summary report with complex filtering and grouping: aging buckets, on-account/unapplied payment filter, financial/accounting periods, etc.- Generated 100k+ invoice and payment data and uploaded through custom integration for performance testing.- Worked with PM and Technical Mentor to define requirements, and presented to organization managers.---------------------Rotation 2: Back EndLanguages and Tools: XpressO (XO, Workday proprietary), Java, Neo4j, Python, Jupyter- Created a tool to extract and analyze a property graph model of Workday’s internal business logic. Business objects and functions were turned into hundreds of thousands of nodes and relationships.- Used tool (with internship project tool) to align stack traces commonly found in thread dumps with path through internal XO object graph, giving us insight into business logic performance hotspots.- Presented tool with Technical Mentor at company-wide conference to accelerate adoption.---------------------Rotation 3: DevOpsLanguages and Tools: Terraform, Helm, Kubernetes, Jupyter, AWS, Bamboo, Prometheus, Grafana, Airflow- Created custom JupyterHub Docker container, then used Terraform to install it on a Kubernetes cluster (via Helm) on AWS.- Modified an open source Bamboo plugin to export build metrics to a Prometheus time series database.- Built SRE reports with Grafana and automatically published them each week in Jupyter notebooks with Airflow Papermill operator.

Languages and Tools Utilized: Java, ANTLR, Git (Bitbucket), JIRA- Developed a state-machine-driven parser to collect and output statistics from thread dumps to a database.- Created a grammar and AST visitor using ANTLR to incorporate a user-specified grammar and options file.- Worked closely with technical mentor to define customer requirements, revise software design, and present.

Languages and Tools Utilized: Visual Studios, C# and .NET, Python, keras, OpenCV- Learned the company’s software system, then fixed bugs, added GUI features, and modified the state machine.- Used machine learning and machine vision techniques to find and read datamatrix id’s from poor-quality pictures of wafers. Achieved over 50% accuracy on industry leader’s successes without using wafers’ alphanumeric id’s.- Experimented with neural networks, adaptive thresholding, template matching, etc. and delivered weekly reports.