Case studies / Recommendations
Recommendations
Feeds, home screens and playlists: picking what each person sees next.
Approach
21 write-ups
Tag:
-
ShopifyNewRecommendations
Shopify’s generative discovery model
Problem Personalizing search and recommendations across billions of products while staying within live request latency.
Solution Trained a shared generative model on Shopify's search and recommendation data, adapting a pretrained language model with commerce signals to generate product candidates in real time.
Read on shopify.engineeringDetails -
MetaNewRecommendations
Your Algorithm: Giving People More Ways to Personalize Facebook Recommendations
Problem Users could not quickly tell Facebook what they wanted to see or avoid in their feed.
Solution Created a system where people type preferences, LLMs store and track them, then retrieve matching posts via search, topic clusters, graph links, and re‑rank them for feed and reels.
Read on engineering.fb.comDetails -
MetaRecommendations
GenRecs Tokenization: Transforming Ads Recommendation at Meta with Stable Model Vocabulary
Problem Ads struggled to recommend brand‑new products because they had no past click data.
Solution Created GenRecs Tokenization, turning each ad, product, or page into a short sequence of semantic tokens using LLM embeddings, so the model can reason about unseen items like familiar ones.
Research behind this, on AI Papers DailyUtility-Aware Generative Recommendation: Utility-Enriched Semantic Codes and Utility-Guided DecodingRead on engineering.fb.comDetails -
SpotifyRecommendations
Advancing Generative Recommendation Inference: our collaboration with NVIDIA on TensorRT LLM
Problem Personalized recommendations needed real‑time, multi‑surface inference that could scale beyond a proof‑of‑concept.
Solution Spotify built a generative recommendation system that uses a large language model with beam search to generate semantic IDs, then switched from vLLM to NVIDIA TensorRT LLM for GPU‑optimized inference.
Research behind this, on AI Papers DailyTAMEing the Open-World Personalization: Towards Open-Set Personalized MLLM AssistantRead on engineering.atspotify.comDetails -
NetflixRecommendations
Measuring the Impact of Personalized Recommendations
Problem How much value does personalized recommendation add to user viewing?
Solution They built a behavioral model that separates a member’s underlying preference from the extra effect of a recommendation, allowing causal measurement of recommendation impact.
Read on netflixtechblog.comDetails -
NetflixRecommendations
Integrating Netflix’s Foundation Model into Personalization applications
Problem Multiple specialized models made personalization hard to maintain and update.
Solution Created a single foundation model and integrated it into production using embeddings, subgraph queries, and daily fine‑tuning to serve different latency and stack needs.
Read on netflixtechblog.comDetails -
NVIDIARecommendations
Deploying an HSTU Generative Recommender with NVIDIA Dynamo-Triton
Problem Personalized recommendations were slow because models had to process long user histories and large item tables.
Solution They built a generative recommender served with Dynamo‑Triton, using ahead‑of‑time compiled PyTorch, GPU‑backed KV caching, and an embedding cache to cut latency.
Up to 5.93× lower latency
Research behind this, on AI Papers DailyUtility-Aware Generative Recommendation: Utility-Enriched Semantic Codes and Utility-Guided DecodingRead on developer.nvidia.comDetails -
AirbnbRecommendations
Personalization without user identity
Problem New visitors arrive without login or history, so the system cannot personalize search and recommendations.
Solution They created Proximity Features that group nearby anonymous users and use the group’s past behavior as signals, enabling instant, location‑based personalization without personal IDs.
Read on medium.comDetails -
AirbnbRecommendations
The guest journey, updated in real time: extending Airbnb’s sequence recommender with Chronon
Problem Guest activity from the same day wasn’t reflected in rankings, making recommendations stale.
Solution Added near‑real‑time model transforms and push mode to Chronon, instantly updating guest embeddings as they browse, so the ranking model sees fresh data.
Research behind this, on AI Papers DailyUtility-Aware Generative Recommendation: Utility-Enriched Semantic Codes and Utility-Guided DecodingRead on medium.comDetails -
YelpRecommendations
ML based ranking using Nrtsearch
Problem Search relevance suffered because traditional term‑based scoring couldn't use rich signals or learn feature weights.
Solution They added an Inference Plugin to Nrtsearch that runs ML models like XGBoost inside the search engine, extracting features from documents and re‑ranking results on‑the‑fly.
Read on engineeringblog.yelp.comDetails -
RedditRecommendations
SnooRec: A Shared Retrieval Foundation for Reddit Recommendations
Problem Different product surfaces needed personalized recommendations but each optimized for different user actions.
Solution Reddit built SnooRec, a shared retrieval system that uses task‑specific user vectors, a common item index, and a control layer to mix intents and query a single ANN index.
Read on reddit.comDetails -
SalesforceRecommendations
Engineering Real-Time Mobile Personalization Across iOS, Android, React Native, and Flutter
Problem Need consistent real‑time personalization on iOS, Android, React Native and Flutter without breaking layouts or user identities.
Solution Created a unified mobile personalization platform with shared backend services and lightweight SDKs for each stack, keeping layouts intact and identities synced.
Read on engineering.salesforce.comDetails -
FlipkartRecommendations
LLMs for Relevance: Automating High-Quality Product Relevance Labeling in Flipkart Search
Problem Manual relevance labeling was slow, expensive, and noisy.
Solution They trained an LLM to judge query‑product relevance, first with supervised fine‑tuning and then with Grouped Relative Policy Optimization to improve consistency.
Matches or beats human annotators at a fraction of the cost
Read on blog.flipkart.techDetails -
EtsyRecommendations
Kafka App? There’s a Skill for That
Problem Shoppers needed up‑to‑date recommendations, but data lagged behind their browsing activity.
Solution Implemented a Kafka streaming pipeline that continuously updates feature vectors and embeddings, feeding them into the recommendation engine in near real‑time.
Read on etsy.comDetails -
AirbnbRecommendations
Personalizing Airbnb search by learning from the guest journey
Problem Hand‑crafted search features couldn't keep up with growing data and limited personalization.
Solution Created a Transformer‑based sequence model that encodes guests' view and booking events, using caps and sampling to handle huge event logs and improve ranking.
Research behind this, on AI Papers DailyThink Globally, Generate Personally: Multi-View Thinking with Adaptive Generation for Generative RecommendationRead on medium.comDetails -
InstacartRecommendations
Semantic IDs: Product Understanding at Scale
Problem Products in different categories never appear together, hurting new and niche items.
Solution Created semantic IDs by turning product embeddings into short integer codes with a residual vector quantizer, letting similar items share prefixes and be discovered across categories.
Research behind this, on AI Papers DailyUtility-Aware Generative Recommendation: Utility-Enriched Semantic Codes and Utility-Guided DecodingRead on tech.instacart.comDetails -
InstacartRecommendations
Scaling Personalized Marketing for Multi-Tenant Commerce Platforms
Problem Retailers needed personalized marketing that kept their brand separate from Instacart’s system
Solution Instacart added a per‑retailer workspace in a third‑party marketing platform, automating provisioning and isolating data, templates, and settings
Read on tech.instacart.comDetails -
MyntraRecommendations
From Products to Inspiration: Inside the Engine of Occasion-based outfit visualiser
Problem Shoppers struggled to picture how clothes fit into real-life occasions, causing many to leave the site.
Solution Created a visual outfit feature that maps items to style recipes, composites images, and uses an LLM‑driven diffusion model to generate realistic scene pictures.
Read on medium.comDetails -
InstacartRecommendations
Our Early Journey to Transform Instacart’s Discovery Recommendations with LLMs
Problem Static content libraries limited personalized discovery for shoppers.
Solution Created an AI‑native platform that uses LLMs to generate, evaluate, and retrieve content, delivering more relevant and cohesive recommendations on the Shopping Hub.
Read on tech.instacart.comDetails -
ShopifyRecommendations
The generative recommender behind Shopify's commerce engine
Problem Recommendations used simplified signals and missed the order of shopper actions, leading to less relevant product suggestions.
Solution Built a generative recommender that reads full buyer event sequences and predicts the next product with an autoregressive model, serving in real‑time at Shopify’s scale.
Research behind this, on AI Papers DailyThink Globally, Generate Personally: Multi-View Thinking with Adaptive Generation for Generative RecommendationRead on shopify.engineeringDetails -
TripadvisorRecommendations
Travel Plus Experience Meets Data at Viator
Problem Travelers saw hundreds of irrelevant options, making it hard to pick the right activity.
Solution Created data pipelines and attribution models that rank and filter results, showing only the most relevant, personalized experiences.
Read on medium.comDetails
No write-ups match these filters.