Blog
Recommendation systems, AI discovery, and ranking infrastructure
Technical guides on AI-powered personalisation, collaborative filtering, product discovery, evaluation, and production recommendation systems.
Technical foundations
ALS collaborative filtering and cold-start recommendations
Start with the core collaborative filtering model, then follow the path into cold-start handling and offline evaluation.
Read the ALS guide →
AI product discovery
AI shopping assistants and ecommerce product discovery
Track how assistant-led shopping changes discovery, ranking, and the data quality retailers need before the click.
Read the product discovery piece →
Recommendation systems hub
Recommendation systems architecture, ranking, and metrics
Use the cluster page for the broader route through ranking, retrieval, signal quality, governance, and measurement.
Open the recommendation systems cluster →
Start here
Core recommendation systems paths
A shorter route into the posts Google is already testing: ecommerce product discovery, ALS collaborative filtering, and the current recommendation systems landscape.
AI shopping assistants are becoming a brand discovery layer for ecommerce product discovery. That changes where relevance gets decided, who controls the shopping journey, and how recommendation systems need to adapt.
A technical walkthrough of ALS collaborative filtering, Alternating Least Squares, weighted implicit feedback, and the Hu-Koren-Volinsky framework behind production recommendations.
From mandatory data quality frameworks to the collapse of standalone vector databases, here's how recommendation systems are evolving in 2026 and what it means for your platform.
Archive
Latest writing
Browse the full archive across recommendation quality, ranking systems, AI agents, embeddings, infrastructure, and product discovery.
AI agents are becoming useful operators for recommendation systems when they can inspect contexts, search catalogues, explain rankings, and record events through governed tools.
Recent research reveals a critical gap between offline evaluation metrics and real-world recommendation performance. Here's what engineering teams need to know about building systems that actually drive business outcomes.
Learn to boost, pin, filter, and diversify your recommendation results using NeuronSearchLab's rules engine. A complete guide to fine-tuning algorithmic outputs with business logic.
A step-by-step guide to building a TikTok-style For You feed algorithm on NSL: short-form video events, catalogue ingest, ranking pipelines, training, and attribution.
NeuronSearchLab now supports a broader retrieval and ranking stack: content embeddings, two-tower retrieval, gSASRec-style sequential retrieval, Semantic-ID generative retrieval, XGBoost rankers, MMoE, PLE, and bounded reranking.
From LinkedIn's feed redesign to Meta's LLM-scale ranking models, large language models are fundamentally changing how recommendation systems work. Here's what the latest research and production deployments tell us about the future.
As LLMs reshape recommendation architecture and AI agents filter marketplace discovery, the challenge shifts from building better algorithms to governing them responsibly. Here's what the latest developments reveal about the future of recommendation governance.
Most recommendation systems fail silently when training and serving data distributions diverge. Here's how to detect drift before it destroys ranking quality.
Meta's LLM integration signals a fundamental shift in recommendation systems. Here's what the convergence of language models and personalization means for platform builders.
Most teams track CTR and call it done. Here's what NDCG, hit rate, and catalogue coverage actually measure — and how to use them before you ship.
Every recommendation system has a cold-start problem. This post explains practical production strategies for user and item cold-start paths.
Most recommendation vendors give you a black box. We built an open evaluation harness so you can score any system — including ours — with your own data.
Recent updates from Meta, Google, retail media infrastructure providers, and agentic commerce platforms point to the same lesson: recommendation quality increasingly depends on richer feedback, stronger metadata, and production-grade retrieval rather than clicks alone.
Recent platform changes suggest that search, recommendations, and advertising are no longer separate optimisation problems. They are increasingly drawing on the same retrieval, ranking, and feedback infrastructure.
Meta just announced four new chips purpose-built for ranking and recommendations. The strategic lesson applies well beyond billion-user scale.
Recent recommendation-system research suggests that quality is no longer just a ranking problem. Reliability, reproducibility, and safety are becoming part of the evaluation baseline too.
X’s latest transparency push highlights an uncomfortable truth: recommendation quality and recommendation trust are related, but they are not the same thing.
What X’s engineering write-up revealed about candidate sourcing, ranking, and filtering - and what product teams should learn from it.
A practical look at what algorithmic feeds on X reveal about recommendation systems, ranking tradeoffs, and why operator control matters.
What the proposed TikTok-Oracle arrangement revealed about recommendation engines: they are not just product features, they are strategic infrastructure.
Why overly precise recommendation models can harm user satisfaction and what to do instead.
A practical look at how app recommendation engines power user engagement, discovery, retention, and personalised experiences.
A plain-language explanation of how vectors and embeddings power modern ML recommendation systems.
A new kind of recommendation engine that gives you control, flexibility, and lightning-fast personalisation powered by machine learning.