Our Early Journey to Transform Instacart’s Discovery Recommendations with LLMs

Moein Hasani26 Feb 2026

Problem

Instacart's Shopping Hub, the page a customer lands on after choosing a retailer, is built from placements: titled groups of products. Placements are defined by hand with generic titles such as "Dairy", and every user draws from the same library. Separate teams build them, so the page feels chaotic and users scroll without a clear path. Hand-made content is slow to produce and check, which blocks personalization.

Example from the article: a thematic "Breakfast" placement should show waffles and pancakes to one user and granola and yogurt to another.

Solution in brief

A cascade of LLM steps designs the page per user. A page-design agent writes personalized themes, a second model maps each theme to retrieval keywords, and filters remove duplicates and weak or off-brand placements. Existing ranking services then order the finalized products and placements. A three-part evaluation suite, ending in a small fine-tuned classifier, keeps quality checks affordable.

Questions this case answers
  • How to replace a hand-curated, one-size-fits-all content library with generated, per-user content blocks that stay coherent as a page and compatible with an existing ranking stack
  • How to keep the cost of a multi-step LLM generation pipeline low when one step must choose from a very large closed vocabulary
  • How to quality-check generated content at a scale where running an LLM judge on every item is unaffordable
  • How to generate personalized, cohesive placements for a discovery page with an LLM instead of a hand-built content library?

Tools and methods

  • Page design agent (LLM) — turns user context into themes, with a structured output schema.
  • Teacher-student fine-tuning — a closed-weight LLM labels data, an LLM judge prunes it, a smaller model learns from it.
  • Retrieval-augmented keyword selection — embedding neighbors shrink the keyword list in the prompt.
  • LLM-as-a-judge with human validation — page, placement and product-level grading.
  • DeBERTa cross-encoder classifier — product-to-theme relevance on every generated placement.
Continue reading — it's free

Google or a one-time email code. No password.