Post-training Gemini for Search

Working with LLMs (pre-training, preference alignment, and instruction tuning on 10B+ param models) to create smarter agents for the enterprise

Temporarily moved to Paris to learn French.

Lead a team of five responsible for building and deploying the deep neural network models, ensemble-based models, text processing system, model training platform, and label management flow underlying the ML systems. • Led the design, training, and roll-out of many iterations of our deep multi-modal (e.g. text + numerical + behavioral features) model. Utilized the latest in language model techniques (e.g. BERT) to further drive performance of our deep neural network model. Improved on feature representation via different encoding schemes and model architectures. • Proposed, designed, and led the roll-out of the model auto-retraining system that allowed an engineer to automatically train and launch models to live traffic on a specified cadence. This required standardizing much of the ad-hoc modeling workflows into internal libraries while adding automation around evaluating and launching models. The system decreased human time to train models from many hours to less than one hour and maintained recall.• Designed the system to organize and aggregate the output from many models (e.g. GBDT models, language models, neural net models) to maximize system performance for multiple objectives (e.g. spam detection, impersonation detection, phishing detection). This allowed the team toscale out and perform modeling improvements in parallel, increasing the number of models safely in production from 5 to >40• Implemented and rolled out an index-based text matching system for intents and key phrases. • Organized and ran planning quarterly for the team, regularly performed stakeholder management with management and TL's of other teams, performed design and code reviews for my team, and ensured that overall the team was functioning well and able to properly meet our company wide goals• Progressed in role twice, from Machine Learning Engineer ‑> Senior MLE (2021) ‑> TL (2022)

• Machine learning for Video Ads. Developed models at large scale for app installs and in-person store visit conversions. Worked with linear (logistic and Poisson regression) and deep models (Deep Cross Network) to support a billion+ clicks a day• Optimization for Google Cloud Infra. Modeled the resources of data centers and optimized upgrades to the existing fleets of servers using mixed integer programming techniques.