Duolingo's notification algorithm isn't just A/B testing with extra steps. It's a bandit that knows when to shut up — and that's the hard part.

Every recommendation system that works well is fusing multiple signal types. The ones that don't understand this ship vibes-based retrieval and wonder why users leave. A taxonomy of signals, how they combine, and what the SOTA ecosystem gets right and wrong.

Your retrieval pipeline returns results. But does it return the right results? How New Computer used LangSmith's experiment framework to achieve 50% higher recall and 40% higher precision in agentic memory retrieval — and what you can steal from their approach.

Gradient-boosted decision trees for people who've never trained one — how they work, when they win, when they don't, and why tabular foundation models are about to make this conversation more complicated.

The architecture behind every recommendation system that actually works at scale — why splitting the model in half is the key to serving billions of candidates in milliseconds.

Cross-modal pattern detection across behavioral data, free-text, and metadata — the capability that actually justifies using an LLM in a recommendation system.

The same metric across 7, 30, and 90 days tells you where someone is heading, not just where they are. Here's why that distinction is the whole game.

Start with one function call. Add skills when the prompt gets too long. A no-framework guide to building agents that actually ship.

Not all behavioral signals deserve the same compute budget. Genre affinity changes over weeks; session mood changes in seconds. Classify by stability, infer by tier, and stop pretending daily batch is the answer to everything.

Two users search 'sleep' and get different results — with no LLM at query time. How pre-computed, theme-keyed expansion terms turn a flat search into something that actually knows you.

Why your personalization agent should never run at request time. The LLM does its heavy lifting on a schedule; your product serves the artifacts. Zero latency, infinite scale.

What users don't do matters more than what they do. Most recommendation systems are built entirely on applause. Here's why the silence is louder.

One LLM call, multiple output shapes for multiple consumers. Design your schema like a protocol, not an afterthought.

What you say you'll do vs what you actually do — the gap is the insight.

How to avoid turning personalization into an algorithmic echo chamber — and why the fix isn't as simple as randomly throwing garbage at your users.

"After you close the laptop" vs "Your Evening Wind-Down" — why the difference matters, and how to build systems that infer emotion without weaponising it.

Databricks trained an RL-based search agent on GLM 4.5 Air that beats Claude 4.6 and GPT 5.2 on enterprise knowledge retrieval — at a fraction of the cost.