Tech Lead for Roblox AI assistant and code assist

Developed various applied machine learning products, and grew them organically

to 400k monthly active users.

• Presto Trip: A travel planner / exploratory travel search engine that used machine

learning and constraint programming to optimize travel fares.

Wrote all the code for the model and optimization engine in Python. Crawled and

trained a flight pricing model on 70M data points. Used Keras for the model and OR-Tools for the constraint solver.

• ML powered extensions for Google Meet such as background blur, face swap and

AR filters: These extensions execute Tensorflow JS machine learning models in the

browser, and made use of real time face mesh triangulation / post processing to let

users apply video effects during Google Meet calls. Worked on optimizing the

performance to run vision models in a browser in real-time.

• Meme Fusion: A meme maker community aimed at students. Used ML for

features for face replacement and content moderation. Wrote server code for a

newsfeed and live chat. Grew the app to 20k daily users organically.

• User Phone Reviews: A NLP based comparison shopping app that analyzed reviews.

Crawled reviews from various online shopping sites. Wrote and used

semi-supervised labeling to label ~100k sentences and train a BERT model to

classify what phone features a review was focused on.

ML and NLP engineer.

Responsibilities spanned data pipelines, tooling, language models, evaluation pipelines,

setting up human evaluation processes and inference server development.

• Lead a cross functional team of 5 to launch Smart Reply for support agents with a

click through rate of ~25% in production with GSuite support agents.

• Carried out various experiments with data cleaning approaches, input features, and

inference configurations to make language models help solve tech support cases.

• Wrote detailed human rater guidelines used by a team of a dozen cloud support

agents to rate predictions by Smart Reply models.

• Primary owner of Hangouts chat data pipeline at Google.

• This pipeline took 3-5 days to run across ~10k nodes and often failed.

• I undertook a large-scale refactor (~40k lines of C++ code ) of this pipeline which

made it run in ~8 hours and made it modular such that multiple other teams could

partner with us to train Smart Reply models, and easily extend this pipeline with

evaluation metrics and training features.

• The end result was a stable, more testable, fault-tolerant, and extensible pipeline

that ended up being used by over a dozen different teams at Google.

• Lead the launch of internationalized Smart Reply in various languages.

• Was previously also an engineer in Google Apps.

• Worked closely with the many cross functional teams to define and execute our technology strategy on many fronts, including client performance, platform integration, feature planning and development, automated testing, cheat detection, server availability, in app advertising, security, productivity tools and content tools.

• Won the highest individual award at the company for technical leadership.

• Developed tools, practices, and teams that made Cityville a well oiled machine responsible for generating over a million dollars of revenue a day at peak.

• Grew the engineering team and development practices from an small pre launch team to a team of 50+ engineers shipping 3 software releases a day, 5 days a week.

Some projects I was directly involved with or lead:

• Reduced the load time of the game by 60% through a combination of algorithmic and networking optimizations

• Scaled the game infrastructure to let Cityville handle over 20 million users a day

• Zcon: A fast and reliable massive scale software release tool capable of controlling software deployments to many thousands of nodes

• Zops dashboard: A unified reporting dashboard for all services across Zynga, including logging, business analytics, operational alerting and performance metrics.

• Operational expsense reduction projects including migrating Cityville off of AWS to Zynga's private cloud and many other server efficiency projects

• Won a few hackathons. My day job didn't involve coding anymore, so I had to get my fix somewhere.