Vinesh Gudla

Vinesh Gudla is a machine learning engineer at Ambience Healthcare and a former staff machine learning engineer at Instacart, where he helped build the systems that interpret grocery searches and recommend relevant products. His work spans personalized ranking, database infrastructure, language-model-powered query understanding, and the economics of agentic search.

Gudla studied at Stanford University, led search work at Amazon and Groupon, and held founding machine learning engineering roles at startups working on enterprise anomaly detection and data management. At Instacart, he progressed from personalized decision-making to search retrieval and generative AI: in 2023, he co-authored work on contextual bandits for large action spaces; by 2025, his work included production search infrastructure and an LLM-powered intent engine.

• Marketplace-grounded query understanding: General-purpose models can misread commercial intent: an Instacart shopper searching for protein may want bars or shakes, not chicken or tofu. Gudla grounds category predictions in product taxonomies and categories that actually convert, improving retrieval for uncommon queries. His AI Engineer conference talk also describes substitute, broader, and synonymous rewrites that recover results when retailer catalogs differ.

• Hybrid search on Postgres and pgvector: Gudla co-authored Instacart’s account of combining keyword and embedding-based retrieval in Postgres, integrating inventory-aware filtering while simplifying indexing and infrastructure. A production experiment reduced searches returning no results by 6 percent.

• Coherent, cost-conscious search systems: Gudla uses offline generation and caching for frequent queries while exploring distilled models for the long tail. He has also highlighted failures created by disconnected models: a brand detector can correctly recognize a kombucha brand while a spellchecker changes the same query into an unrelated food.

• Adaptive agentic search: In research he described publicly, a lightweight router determines when difficult queries warrant additional reasoning, reducing token costs by approximately 30 percent. His later co-authored research also examines related-intent generation for product discovery and generative storefront recommendations.