Jonathan Gu
Machine Learning Engineer at Coframe and a PhD economist. I build the backend and the models that decide what each visitor sees on a client’s site, including search and recommendations, and I run the A/B tests that measure whether they help. My work covers metric design, data, training, serving, rollout, and experiment analysis.
- Personalization on demand: build the backend and the models that decide what each visitor sees on a client’s store, including search, recommendations, and the order of products on collection pages.
- Search for One Kings Lane: built instant results from precomputed tables, BM25 word search with typo tolerance, embedding retrieval for queries with no word match, a LightGBM ranker trained on expected revenue and blended with best sellers, and rules for stock, price, product family and category. It runs as a live A/B test against the store’s existing search.
- Ranking for each shopper: train models on each store’s own visitors that use the campaign that brought the shopper in, their searches, the products they viewed and their cart. Purchases are weighted by revenue with a cap.
- Serving: a Rust engine on Cloudflare Workers that serves precomputed tables with a lookup and a sort, shopper history in Durable Objects, and deadlines on lookups so the store shows its own ordering if our answer is late.
- From request to model: agents handle much of the work. Each model has a simple baseline to beat, a sample check within ten minutes before full training on Vertex AI, evaluation on held-out shoppers, a preview on the client’s page, approval by a person, and a live A/B test on revenue per visitor. Our best recent internal drill took 9.5 hours from request to serving model. The goal is three hours.
- Uplift Growth Targeting: built the MVP for growth targeting using causal experiment outcomes to prioritize users by expected incremental treatment effect rather than naive conversion propensity.
- Causal Marketing Measurement: measured coupon and marketing incrementality, connecting experiment results, budget allocation, campaign diagnosis, LTV/CAC tradeoffs, and in-app decisioning.
- Ads Optimized Bidding: built Instacart’s first optimized ads automated bidding system from scratch, replacing manual product-keyword bids with weekly-budget bidding across roughly 70% of Sponsored Products spend (~$700M annualized).
- Inventory Intelligence: built item-store availability prediction systems across noisy catalog, store, fulfillment, and shopper signals, including real-time LightGBM serving, calibration, thresholding, rollout, and critical-item prioritization.
- Studied Bing search-auction counterfactuals and revenue effects under alternative ranking rules with Prof. Susan Athey; contributed to research on auction design, ranking rules, and causal effects in marketplaces.
Issued: US 12,511,677 B2, Automated Policy Function Adjustment Using Reinforcement Learning Algorithm (reinforcement-learning control for automated bidding). Additional portfolio: causal targeting, dynamic offers, double-wide ad auctions, retailer classification, fulfillment routing, replacement agents, and item display decisions.
ML & Methods: search and retrieval (BM25, embeddings), learning to rank, recommendations, experimentation, causal inference, uplift modeling, ads bidding, reinforcement learning, auction design
Engineering: Python, SQL, TypeScript, Go · LightGBM, TensorFlow · Airflow, DBT, Snowflake, Kafka, Flink SQL, MLflow, Datadog, Arize
Systems: edge serving (Cloudflare Workers, Durable Objects), training on Vertex AI, real-time serving, feature pipelines, monitoring, rollout