Nice to meet you,
I'm Calvin.

Overview

On the Content Suitability team, I built a proof of concept using foundation models to classify experiences (ie social hangouts) to support age-appropriate surfacing and protect minors from inappropriate content. The approach extended to additional policy dimensions including blood and violence, then was handed off for full production implementation.

I designed and shipped auto-sequestration, an automated system to detect and apply consequences when experiences routinely violate content policy, replacing a workflow of individual moderator takedowns and one-off creator messages. It eliminated on the order of ~180 manual takedowns per week and saved roughly ~90 moderator hours weekly, with a fully automated creator notification flow.

Highlights

• Foundation-model POC for experience classification and age-gating; expanded to blood and violence categories; handed off for production.
• Auto-sequestration at scale: automated detection, consequences, and end-to-end creator notifications.
• Championed agentic development across the team-structured agent configuration, Cursor skills, and workflow automation (including PR creation) to speed up repetitive engineering work.

Overview

I joined CLEAR as the founding full-time engineer on the Subscriptions & Payments team, where I helped scale our operations from initial concept and design through production deployment and team expansion.

I helped architect and implement a custom Stripe integration that revolutionized our payment infrastructure, enabling advanced capabilities like dynamic repricing and flexible trial periods that previously involved long downtime or development cycles.

The new system is projected to recover up to $13 million in revenue from failed payments and subscription renewals. I led the strategic rollout of our mobile client deployment, orchestrating a careful ramp-up from 0 to 100% of production capacity, ensuring a seamless transition while maintaining system stability.

Dynamic re-pricing decreased development effort & downtime required by the legacy system from several sprints to minutes via admin self-service.

First-in-class architecture meant that PII (personally-identifiable information) like PANs don't flow through the new system ensuring security and compliance.

Subscription group upgrading and downgrading functionality allowed users to modify their subscription in turn increasing revenue.

Overview

Millions of users enter hourly time during their shift every week using time management software. While most of the data entered reflects the actual worked time, users can make mistakes. Once time is submitted, managers and timekeepers need to spend considerable time verifying user input before the time is processed for payroll.

As part of the Time and Scheduling Hub for managers, Workday Time Anomalies automatically reveals possible time-entry errors and alerts managers about unusual time entries. The goal is to save time and improve payroll and labor cost accuracy.

At Workday, I built out components of the cross-team Time Anomalies project that allows managers to detect these unusual time entries using machine learning. I also worked to enhance the delivery of this information to users with the Time Approval Summary filters, which allows managers and supervisors to easily discern and sort through discrepancies with their worker's time.

I deployed the feature through an early adopter phase of 10 customer opt-ins that ultimately scaled to general availability for all Time Tracking customers in Workday 2021.

Overview

Create a natural language processing model that matched user resumes with jobs and delivered a daily ranked list of prospective job postings.

Lessons Learned / Challenges

• Balancing performance and functionality: The job scraping and matching for users involved compute-intensive operations due to the ML model. Using serverless asynchronous data processing could not only alleviate server load but could also improve user experience.
• Optimizing the database: A NoSQL database would allow us to quickly update new jobs with flexibility to change our schema as we prototyped.
• Tuning the algorithm: We could stack programmatic rules with the machine learning algorithm to deliver more accurate predictions.

Details

This project was inspired by a friend who came to me with a problem. Recently laid off from a large Seattle-area tech company, he was looking for a new job. While there were many services to help find jobs, finding a good match with unique skills and experiences required many hours manually parsing job descriptions for a good fit.

What if there was a service that could process and match your resume with related jobs and deliver a ranked list of results daily to aid your job search?

With that idea in mind, we worked together to create Job Matcher, a service that scrapes job posting websites and uses natural language processing to compute similarities between the postings and user resumes. It would prioritize skills (for example, Ruby or AWS Sagemaker) and return to the user a set of matched jobs ranked by the machine learning model's scoring criteria daily.

Architecture

The application consists of several main components: The flask server handles the client interaction including user identification and metadata persistence (resume, email, etc.). To communicate with the MongoDB cluster, we exposed pathways via FastAPI to allow for new user creation, updating of user details, and getting the latest matched resumes.