Gautam Kedia

Hello! I’m GK. I live in the SF bay area with my wife & 2 kids. I’m passionate about machine learning (ML), investing & doom scrolling. If you’re building a machine learning product, I would love to hear from you!

Work

2023-Current: Applied ML, Stripe

In 2023, I started a new Applied ML team at Stripe to bring LLMs & foundational models to Stripe. We are exploring applications in our Documentation, Support, and even payment network products.

2020-2023: Fraud Intelligence, Stripe

I led the Fraud prevention ML & engineering teams at Stripe. We reduced losses by double digits YoY while improving onboaring experience for Stripe users. We built state-of-the-art ML and backend systems that scaled with Stripe. Here are some of the models & systems the team shipped:

Transformer model on sequences of API calls to detect fraud patterns
Unsupervised clustering model that finds novel fraud attacks
Large scale distributed fraud detection model
Anomaly detection system to detect risky behavior
Bringing ML infra to industry standards – auto-training, feature store, and debugging
Online streaming system to run ML models and make decisions in realtime

2018-2020: Applied ML, Lyft

I pitched and led the Applied ML team. Here are some projects that the team delivered:

Ride destination recommendations with a transformer based model
Revenue-optimizing mode recommendations model
Short-term demand & supply forecasting with Conv-LSTM model
A Thompson Sampling based adaptive experimentation system at Lyft
Optimized tolls prediction pricing model
Coaching program for drivers who drive with their phone in their hands
Optimized ad spend model for driver acquisition
Prototype for RL based routing algorithm (KDD Cup 2020 3rd place)
Cancel fee forgiveness model
Late ride discount coupon model

2015-2018: Office AI, Microsoft

I led the Smart Reply & Office AI prototyping team. Here are some projects that the team delivered:

State-of-the-art reccurrent neural net for predicting replies to Skype messages & Outlook emails
- ~100M parameter model trained on ~1B emails and serving ~10B emails per day!
- Handled correctness and bias in models
Prototypes for meeting detection, task detection, and summarization of emails

2010-2014: Bing, Microsoft

I led the Location Personalization & Cortana Ranking team at Microsoft.

Consulting

I’m helping Oxylabs develop their ML strategy

gk