Deep Dives
Papers, architectures, and systems explained without the bullshit
How modern rec systems retrieve, rank, and serve content at scale.
WTF is the Recovering Difference Softmax Algorithm?
Duolingo's notification algorithm isn't just A/B testing with extra steps. It's a bandit that knows when to shut up — an…
Signal Fusion: How Semantic, Relational, and Direct Signals Combine to Make Recommendations That Don't Suck
Every recommendation system that works well is fusing multiple signal types. The ones that don't understand this ship vi…
Optimising Recall and Precision in LangSmith Experiments
Your retrieval pipeline returns results. But does it return the right results? How New Computer used LangSmith's experim…
WTF is Two-Tower Recommendation?
The architecture behind every recommendation system that actually works at scale — why splitting the model in half is th…
Signal Stability Classification: Inference Cost-Benefit in Hybrid Recommendation Systems
Not all behavioral signals deserve the same compute budget. Genre affinity changes over weeks; session mood changes in s…
Negative Signals as First-Class Citizens in Recommendation
What users don't do matters more than what they do. Most recommendation systems are built entirely on applause. Here's w…
Filter Bubble Mitigation in Personalized Systems
How to avoid turning personalization into an algorithmic echo chamber — and why the fix isn't as simple as randomly thro…
Understanding users through what they do, not what they say.
Through-Line Detection: What LLMs See That Rule Systems Can't
Cross-modal pattern detection across behavioral data, free-text, and metadata — the capability that actually justifies u…
Multi-Window Temporal Aggregation for Behavioral Trajectory
The same metric across 7, 30, and 90 days tells you where someone is heading, not just where they are. Here's why that d…
Query-Theme-Keyed Search Expansion
Two users search 'sleep' and get different results — with no LLM at query time. How pre-computed, theme-keyed expansion …
Intention-Action Gaps as Behavioral Signals
What you say you'll do vs what you actually do — the gap is the insight.
Empathy Architecture: Designing LLM Outputs That Don't Feel Like Surveillance
"After you close the laptop" vs "Your Evening Wind-Down" — why the difference matters, and how to build systems that inf…
Patterns for building real systems with large language models.
From Single LLM Call to Deep Agent: An Honest Migration Path
Start with one function call. Add skills when the prompt gets too long. A no-framework guide to building agents that act…
Pre-computed Personalization: The Offline Agent Pattern
Why your personalization agent should never run at request time. The LLM does its heavy lifting on a schedule; your prod…
The Multi-Artifact Output Pattern
One LLM call, multiple output shapes for multiple consumers. Design your schema like a protocol, not an afterthought.
KARL: Knowledge Agents via Reinforcement Learning
Databricks trained an RL-based search agent on GLM 4.5 Air that beats Claude 4.6 and GPT 5.2 on enterprise knowledge ret…
The algorithms and architectures underneath everything else.
Everything Else
-
2026-03-17
Three Loops Your LLM Pipeline Needs From Day OneZero-cost architectural decisions — a prompt registry, per-block model config, and three BigQuery columns — that cost nothing to add now and unlock DSPy optimisation, LangSmith experiments, and behavioural flywheels later.8 min
-
2026-03-15
Consistency Models for Edge-Native Memory SystemsStrong consistency is a tax you pay for guarantees you don't need. Why AI memory systems are accidentally CRDT-shaped, and what that means for edge-native architecture.15 min