Saboo’s path into machine learning began with chemical-plant automation. Working with feedback systems drew him toward statistics and machine learning; he also explored energy policy through work with the World Bank before entering machine-learning consulting, where he built solutions across industries. He holds a master’s degree from Duke University and combined BEng and MEng degrees from Imperial College London. His career account traces these transitions through his subsequent work at Amazon’s Alexa AI.
At Amazon, Saboo developed first-generation language models and distilled models for Alexa AI’s new language launches. Transferring a powerful teacher model’s capabilities into smaller models suited to deployment remains central to his later work. At DoorDash, he co-authored an account of Ask DoorDash’s intelligence and memory architecture, bringing language understanding together with retrieval, ranking and personalization.
DoorDash’s expansion beyond restaurants creates a specific personalization problem: a customer can have a substantial restaurant-order history and almost no grocery-shopping history. Saboo and Nimesh Sinha were equal contributors to Mind the Gap, a 2025 research project that uses language models to transfer useful preferences across categories. Its hierarchical retrieval-augmented generation pipeline identifies broad product affinities, narrows the relevant categories and then makes more specific predictions. Those predictions become features in an existing multi-task ranker, helping personalize a new category before the customer has accumulated a long purchase history there.
Teaching systems what products and preferences mean
Saboo argues that marketplace discovery requires understanding shopper intent and constraints alongside engagement. A popular product can still be the wrong answer: ordinary spaghetti fails a request for gluten-free pasta, regardless of how well it sells. His collaborative research and engineering work develops several ways to make that understanding usable in production:
Graded relevance: In Mine and Refine, which Saboo co-authored, a lightweight language model learns relevance guidelines from human annotations. An initial retriever finds difficult examples, which are relabeled before a second contrastive training stage sharpens the distinctions between exact matches, acceptable substitutes and merely related products. In ranking, a relevance prediction can sit alongside click, add-to-cart and conversion objectives, allowing the system to balance semantic fit with engagement.
Semantic product IDs: Saboo co-authored One Hierarchy, Two Systems, which learns a product hierarchy from content embeddings. Shared prefixes capture broad relationships, while later parts of each code distinguish products more finely. Ranking systems can share evidence across related products, including new or rarely purchased items; search can use the same hierarchy to refine queries. Constraining suggestions to the merchant’s assortment keeps semantic relationships connected to products the customer can actually buy.
Unified consumer memory: Saboo and his collaborators developed memory blocks that express preferences in language and provide embeddings and graph features to conventional models. The design separates historical behavior, immediate session context and explicitly stated preferences. These representations let shopping agents and ranking systems reuse customer understanding without each rebuilding it for its own task.
Offline generation, online personalization: Saboo co-authored DoorDash’s work on personalized product carousels, extending personalization to how a store page organizes and describes products. Language models generate titles, subtitles and shopping intents from consumer memory in batch. When a customer opens a store, retrieval and category lookup resolve those intents into products, which existing systems merge and rank. This supports individually generated merchandising without a fresh language-model call for every page load.
His approach to search and personalization places expensive language-model reasoning offline, then transfers it into reusable labels, catalog representations and customer context. Smaller models and established serving systems put that understanding to work across retrieval, ranking and generated collections.
DoorDash uses expensive reasoning offline to teach fast retrieval and ranking models what shoppers mean. Graded relevance labels, semantic IDs and reusable consumer memory then support everything from better search results to personalized store collections.
Graded relevance separates satisfying a shopper’s constraints from merely being popular or related. DoorDash uses offline LLM labels to teach that distinction to both retrieval and ranking.
Two-stage retrieval training uses the first model’s mistakes to build a harder curriculum. Ranking adds an explicit relevance objective alongside clicks, add-to-cart events and conversion.
Semantic IDs and consumer memory supply reusable meaning: catalog neighborhoods help with sparse items and query reformulation, while text, vectors and graphs make shopper context useful to different model types.
Personalized collections combine offline LLM generation with existing item hydration and ranking. The broader architecture captures reasoning once and lets several fast serving systems benefit from it.
▶ Watch ↗