Building LLM-based applications to create a GenAI-powered user experience for all YouTube users.

Worked as an L4 Mid-Level SWE on ML Infra in the Account and Behavioral Abuse (ABA) team at YouTube, focused on creating and maintaining the core architecture that protects YouTube’s ecosystem from manipulation and circumvention by adversarial actors

Fully owned and drove a multi-quarter effort to launch and land a notification-based pipeline for triggering ML model classification at feature readiness, reducing model enforcement latency from 20m@P50 to <1m@P50, increasing model robustness, reducing redundant model triggers, and increasing infra flexibility

Introduced a new ML Infra component, leveraging UDFs in the Spanner DB, that bypassed existing SQL limitations and unlocked a set of new model feature computations (such as quantile features), while reducing feature dev velocity from ~1mo to <1wk for existing feature generation/aggregation

Extended the triggering mechanism to support feature generation for graph-based ML models, unblocking launches for the first ML-based MAC detection classifier (auto-terminating MAC violative channels at >90% TVC precision) and a supervised channel-to-channel graph model (contributing ~19bps to a top-level TnS metric)

Designed an automated latency monitoring system to reduce feature onboarding dev velocity by 1wk, while setting up E2E latency monitoring/alerting for 7 new channel feature families

Set up E2E integration testing (using ITS) for the team’s entire low-latency ML infra pipeline, providing coverage over all systems with 4 different functional tests and connecting over 9 different services and infra components

Joined the YT ABA Core oncall rotation in Q2 2024 – ensuring system reliability at <30m response time

Awarded 1 spot bonus (low latency model launch) and 2 peer bonuses (improving onboarding, adding new ML signals)

Joined Google as an L3 New Grad SWE working on ML Infra, getting internally promoted to L4 Mid-Level SWE slightly over a year after joining

Cross-collaborated with the Creator Abuse team to launch ABA’s first low-latency channel spam ML model, reducing time-to-detection for adversarial channels from 40hr@P95 to <80m@P95 and increasing model recall by 97.6% for ‘racy shorts videos’

Contributed to the launch of Plexus Real Time – the first real-time ML model (built in TensorFlow) for engagement abuse playback detection, leading to 298M/day previously unclassified, ML-eligible playbacks being marked, while also increasing model robustness to adversarial reverse engineering

Worked closely with the Actor Understanding team to design and launch a low-latency solution to reduce Regretted Views from terminated channels, leading to 5M+ (previously uncaught) regretted views/day and the suspension of an additional 52K+ channels/day

Designed and implemented per-engagement rule shutoff capabilities for ABA’s Despam Service, which was especially pertinent due to a recent YT Postmortem on rule overmarking

Created new front-end pages and updating existing ones using React, given Figma design mocks and strict deadlines

Introduced the React-Testing-Library (RTL) for React component testing and updated all previously created Jest tests to fit the RTL framework, drastically reducing boilerplate code and standardized the firm’s testing philosophy

Configured and integrated RuboCop (Ruby linter) throughout the entire codebase, introducing Ruby linting into CI workflow

Wrote extensive integration, component, and unit tests using Jest and the RTL

Performed exhaustive pre-deploy quality assurance (QA) checks to smoothly introduce new features into existing products

Participated in scrums, code reviews, and sprint planning in an Agile environment, while communicating directly with project managers and designers across different teams on new feature specifications and implementation details

Developed a Natural Language Processing (NLP) pipeline (preprocessing, model training, analyzing, reinforcing model) using SpaCy for Walmart’s database of driver and customer reviews, resulting in the extraction of meaningful insight and key issues

Created a Convolutional Neural Network (CNN) model using TensorFlow to analyze incoming reviews, distilling them into short problem statements and relevant action items, increasing readability and decreasing time spent perusing reviews

Designed and implemented a custom user interface (UI) to provide a rapid-response one-stop platform for review concerns, decreasing review analysis time by 87.5% alongside providing customers with quicker and better support