Articles

Writing

Technical deep-dives into the architecture of modern machine intelligence, retrieval systems, and neural optimization.

Latest

The Only Moat You Control Is Being Able to Leave

17 Aug 20266 min read

Open weights now land within points of the closed frontier, and the practical consequence is not that you should self-host. It is that the ability to switch has become a design property worth paying for.

AILLMsOpen Source
Nobody Designed the Interface for Watching a Machine Think
13 Aug 20266 min read

Agents now run for minutes or hours, and the interface conventions we have — spinners, chat bubbles, progress bars — were all designed for something else. What you show while it works decides whether people trust the result.

AIAgentsUX
Before You Let an Agent Write Anything, Answer the Blast Radius Question
10 Aug 20266 min read

The jump from an agent that reads to an agent that acts is the largest one in the field, and most teams make it by accident. What matters is not whether it will make a mistake, but what the worst mistake can reach.

AIAgentsSecurity
Most Multi-Agent Systems Are One Agent With Extra Steps
6 Aug 20266 min read

Splitting a problem across several agents is a real technique with a narrow set of cases where it wins. Outside those cases it mostly buys you the same result, slower, with more ways to fail.

AIAgentsArchitecture
Everything Your Agent Learned Today, It Forgets Tonight
3 Aug 20266 min read

Persistent memory is the obvious missing piece in every agent product, and the reason nobody has shipped it well is that remembering is easy and forgetting correctly is not.

AIAgentsMemory
Cost Is a Design Constraint, and Pretending Otherwise Is Why Your AI Feature Got Cancelled
30 Jul 20266 min read

Token spend is not a line item you optimise after launch. It shapes what you are allowed to build, and the teams who treat it as architecture ship features that survive the second invoice.

AICostArchitecture
You Are Not Competing on the Model, You Are Competing on the Harness
26 Jul 20266 min read

The scaffolding around a model — how tools are described, how state is carried, how failure is recovered from — moves real-world performance more than the model swap everyone argues about.

AIAgentsEngineering
Building a Platform for Agents, Not Just Another Agent
22 Jul 20265 min read

Why we stopped building one-off AI agents and built the framework that lets any team build them.

AIAgents
The Gap Between a Magical Demo and a System People Can Trust
21 Jul 20264 min read

Why the last ten percent of making an AI feature reliable in production is harder than the ninety percent that made it impressive.

AIEngineering
Building a Chatbot That Actually Remembers the Conversation
19 Jul 20266 min read

Notes on the problem of building an enterprise chatbot that stays coherent across turns, routes to the right assistant, and answers questions grounded in real business data.

AIChatbot
The Unglamorous Problem That Decides Whether Your AI System Works
16 Jul 20265 min read

Why turning messy, real-world documents into something an AI system can actually use is harder than it looks, and why it matters more than the model you pick.

AIRAG
No, RAG Isn't Dead, You Just Found Out What It Was For
10 Jul 20264 min read

A practitioner's rebuttal to the idea that giant context windows make retrieval obsolete.

AIRAG
Nobody Wants to Build the Evals, and That's Why the Product Breaks
3 Jul 20264 min read

The least glamorous part of shipping an LLM product is measuring whether it actually works, and it's the part that most teams skip until it hurts.

AIEvals
Context Is the Product, Not the Prompt
28 Jun 20264 min read

Why the real work of building with LLMs moved from crafting clever prompts to deciding what information the model gets to see at the moment it answers.

AILLMs