At Reddit, I architected a shared PyTorch two-tower retrieval platform for Ads, supporting distributed GPU training, embedding generation, and real-time vector retrieval for Targeting and Shopping Ads. I scaled Keyword Targeting to general availability, increasing the share of eligible impressions matched to advertiser keywords by 18%.
I improved training and retrieval performance at Reddit, reducing model training time by 25% and P95 retrieval latency by 20%. I also built an LLM-based evaluation agent that reduced manual evaluation effort by 40%.
At Rokt, I developed conversion modeling and Smart Bidding systems, improving CVR model log loss by 5% and reducing advertiser cost per acquisition by 8% in online campaign experiments. At Meta, I worked on Facebook Dating Anti-Scam Detection, increasing scam detection recall by 12% at a fixed false-positive rate.
At Amazon, I developed ML components for Alexa speech recognition and natural language understanding, and built distributed demand forecasting pipelines for Amazon Retail. My work improved slot-level F1 by 3 percentage points and reduced forecasting data preparation time by 22%.

