LinkedIn Feature Trails

Feature trailActive

Semantic Search Stack

LinkedIn has developed a semantic search experience leveraging LLMs to understand natural language queries and infer user goals. This system uses embedding-based retrieval (EBR) on GPUs for candidate generation and a Cross-Encoder Small Language Model (SLM) for ranking. Optimization techniques like score caching and ranking-depth control are employed for efficiency. Relevance is measured using LLM judges aligned with product policies. The EBR model is a dual-tower bi-encoder trained with contrastive loss. This post details the engineering effort behind LinkedIn Premium All-in-One, a unified product for small businesses. It focuses on leveraging AI and LLMs to create unified workflows for sales, marketing, and hiring. Key technical aspects include an LLM-based approach for inferring Target Audiences (TA) using profile and company information, with privacy guardrails. The TA definition is further refined through activity-based signals and implicit signals (buyer-side and seller-side). The system in this post details the engineering effort behind LinkedIn's data context layer, a system that combines automatically generated dataset profiles with team-maintained knowledge about meaning, usage, and limitations to help people and AI agents discover the right assets and understand their context. The system uses an AI enrichment pipeline to generate baseline context from schemas, source code, and query logs, and teams supplement this with human-curated knowledge. The generated layer uses a four-stage pipeline (nomination, collection, generation, persistence) to process signals like query logs, source code, and sample values to create structured dataset profiles and embeddings. The human-curated layer uses a Git repository and pull requests for authoring and review, organizing context into global, team, and analysis scopes.

2 posts · 2026

Timeline

Feature trailActive

Capacity Buffer Optimization

This post details the engineering work to reduce capacity buffers by applying statistical modeling, specifically the newsvendor problem, to better balance hardware efficiency and reliability in a private cloud environment. The approach involves high-fidelity demand signals, strategic aggregation, quantifying organic variability, leveraging resource fungibility, navigating lead-time volatility, and anchoring to service SLAs. This methodology was successfully applied to the Stream-Processing platform, decreasing its buffer by ~8.5% and establishing a scalable methodology for the entire Infrastructure ecosystem.

1 post · 2026

Timeline

Feature trailActive

Semantic Search Stack

This post details the engineering effort behind LinkedIn Premium All-in-One, a unified product for small businesses. It focuses on leveraging AI and LLMs to create unified workflows for sales, marketing, and hiring. Key technical aspects include an LLM-based approach for inferring Target Audiences (TA) using profile and company information, with privacy guardrails. The TA definition is further refined through activity-based signals and implicit signals (buyer-side and seller-side). The system in this post builds upon this foundation by detailing the semantic search system (MUSE) for AI agents, specifically for the Hiring Assistant. It describes the MUSE platform which learns from expert judgments to encode understanding into a model capable of scoring over a billion member profiles in real time. The system uses a dual-tower Siamese architecture with Matryoshka embeddings for efficient retrieval and ranking. It also details the MUSE Teacher, an LLM-as-a-judge system for scalable supervision, and the MUSE Embeddings architecture, including the query and profile towers and the use of 2048-dimensional truncations for ANN search.

4 posts · 2026

Timeline

Feature trailActive

LLM-Based Follows Recommendation System

LinkedIn rebuilt its Follows Recommendation system using LLM-based semantic retrieval and ranking. This new system encodes member and creator profiles into embeddings, enabling topical relevance matching over popularity. It includes offline and online embedding-based retrieval pipelines and integrates these embeddings into downstream ranking models through supervised projection layers. The system has shown significant lifts in follow rates, especially for new members. Further optimizations include parameter-efficient fine-tuning (PEFT) with Low-Rank Adaptation, mixed precision training, and gradient checkpointing for efficiency. The retrieval pipeline uses FAISS for exact KNN search in offline scenarios and a hosted vector search service for low-latency online retrieval. Dimensionality reduction for embeddings in ranking models is achieved via task-aware supervised projection layers.

1 post · 2026

Timeline

Feature trailActive

AI-Assisted Code Review System

LinkedIn has developed a multi-agent AI code review system to address the bottleneck of code reviews at scale. The system uses multiple AI models in parallel, with findings deduplicated and cross-validated to ensure high signal and relevance. A 3-tier customization framework allows for adaptation to specific codebase conventions and organizational policies. The system is operated as critical enterprise infrastructure with a focus on speed, reliability, and observability, achieving a high acceptance rate of 63.9% for automated review bots. This iteration focuses on enhancing the system's adaptability to diverse codebases through a 3-tier customization framework and improving its operational rigor as critical enterprise infrastructure, including staged rollouts, paged on-call, and end-to-end observability.

1 post · 2026

Timeline

Feature trailActive

Generative Recommender Training Efficiency

LinkedIn rebuilt its Follows Recommendation system using LLM-based semantic retrieval and ranking. This new system encodes member and creator profiles into embeddings, enabling topical relevance matching over popularity. It includes offline and online embedding-based retrieval pipelines and integrates these embeddings into downstream ranking models through supervised projection layers. The system has shown significant lifts in follow rates, especially for new members. Further optimizations include engineering techniques to optimize the training efficiency of Generative Recommenders (GRs) at LinkedIn, focusing on data pipeline, compute, and training lifecycle optimizations. Optimizations include a C++-based fused data loader, improved attention kernels (FlashAttention-3, FlexAttention), fused metrics calculation, fused optimizer, and fused embedding table lookups. These optimizations reduced end-to-end GPU hours by up to 65% in internal production training workloads.

2 posts · 2026

Timeline

Feature trailActive

AI-Driven Autonomous Testing

LinkedIn has developed a Quality Assurance (QA) Agent, an AI-driven autonomous testing system that reimagines software quality. This system leverages generative AI and Vision-Language Models (VLMs) to create agents capable of navigating and testing mobile and web applications. The agent uses a hybrid architecture combining deterministic replay (System 1) for stable flows and VLM-based planning (System 2) for dynamic scenarios. Evaluators act as guardrails, and a human-in-the-loop process allows for natural language test authoring and verification. The system has successfully identified bugs and regressions, and aims to democratize quality ownership.

1 post · 2026

Timeline

Feature trailActive

Product Configuration Center

LinkedIn's Product Configuration Center (PCC) has been enhanced to centralize and standardize business configurations across its monetization ecosystem. This evolution addresses the fragmentation and manual management of configurations by introducing a unified platform built on Change Requests, SAGA-based cross-domain orchestration via Change Transport Management (CTM) and Temporal, and deterministic identifier generation. This has significantly accelerated go-to-market velocity, reducing launch cycles from weeks to hours, and improved reliability by eliminating drift and manual errors. The unified graph of configurations also lays the foundation for AI-assisted configuration suggestions and "what-if" analyses.

1 post · 2026

Timeline

Feature trailActive

Internationalization of Agentic AI Systems

This post details the process and challenges of internationalizing LinkedIn's Hiring Assistant, an agentic AI product, for French and German markets. It highlights the difficulties of on-the-fly text generation compared to traditional software localization, focusing on linguistic nuances like gender agreement, formal/informal address, noun capitalization, and number/date formats. The solution involves a shared rubric framework to capture language-specific rules, per-language Low-rank Adaptation (LoRA) adapters, and an adaptive model strategy to reduce serving costs while maintaining native quality. The system now supports French and German, with a roadmap for more languages.

1 post · 2026

Timeline

Feature trailActive

CTV Ads Platform

This post details the engineering effort behind building LinkedIn's Connected TV (CTV) Ads solution. It covers the core design principles, technical frameworks, and infrastructure leveraged to deliver a scalable, reliable, and measurable advertising experience. Key aspects include integrating with CTV publishers via Real-Time Bidding (RTB) through a deal-driven private marketplace, building a supply management system for onboarding and dynamically controlling PMP deals, and adapting supply prior to CTV needs. It also addresses creative specifications, brand suitability, and cross-device identity for professional targeting on CTV.

1 post · 2026

Timeline

Feature trailActive

Unified Hiring Data Integrations Platform

LinkedIn has developed a unified integrations platform to standardize, reconcile, and deliver hiring data at scale. This platform supports both partner-push (BuildIn) and LinkedIn-owned pull/push (BuildOut) integration models, addressing challenges of semantic variability, entity interdependency, and operational scale. Key design principles include security, isolation, observability, unified data contracts, and idempotent, replayable, and evolvable data paths. The architecture features gateway adapters for BuildIn and BuildOut, a centralized integration core, and leverages Temporal for workflow orchestration and Kafka for streaming. This has reduced partner onboarding time by 72%, improved data coverage and completeness, and provides a unified data view for AI systems like LinkedIn's Hiring Assistant.

1 post · 2026

Timeline

Feature trailCooling

Spark Data Pipeline Optimization

This post details the optimization of LinkedIn Sales Navigator's search data manipulation pipeline, transitioning from MapReduce to Spark and significantly tuning Spark jobs. The primary focus is on reducing the total execution time of over 100 DM jobs from 6-7 hours to approximately 3 hours. Key optimization strategies include pruning the job graph by consolidating dependent jobs, identifying and optimizing critical path bottlenecks, mitigating data skewness through repartitioning based on unique search document ID, and adjusting the number of shuffle partitions. Broadcast join optimization is also mentioned as a technique for joining tables of significantly varied sizes.

2 posts · 2023–2025

Timeline

20232026

Feature trailCooling

Scalable Log Storage System

Northguard is a new log storage system designed for improved scalability and operability, replacing Kafka at LinkedIn. It features a data model based on records, segments, ranges, and topics, with log striping for load balancing. Ranges are used instead of indexed partitions to minimize interruption and facilitate stream processing. Metadata is managed by vnodes, which are Raft-backed replicated state machines.

1 post · 2025

Timeline

20252026

Feature trailQuiet since 2023

Feed Recommendation Algorithm Data Sampling

This post details the engineering effort behind refining the LinkedIn member experience by experimenting with data sampling techniques for the feed recommendation algorithm. The work involved handling large-scale datasets (up to 500 million rows) and utilizing big data technologies like Spark and Hadoop to feed these sampled datasets into machine learning algorithms, resulting in business metric gains. The post also touches upon on-call responsibilities for the global feed, including data generation pipelines.

1 post · 2023

Timeline

20232026

Feature trails