Writing

Notes from the build

Articles on Laravel architecture, retrieval-augmented generation, AI agents, small language models, vector search and agentic commerce — from production work, not demos.

RAG9 min

RAG in production: what actually breaks

The demo answers every question. Then real users arrive and recall collapses. Here is where retrieval-augmented generation actually fails, and what to fix first.

28 August 2026Read

Laravel8 min

The Laravel app that survives year two

Most Laravel projects are pleasant for six months and painful by month eighteen. The difference is a handful of structural decisions made early — and they are not the ones people argue about.

14 August 2026Read

AI Agents8 min

Agents that finish the job

An agent that completes 70% of a task is not 70% useful — it is zero useful, because a human has to check all of it. What separates a demo loop from something you can put in front of customers.

30 July 2026Read

SLM7 min

When a small language model beats a frontier model

Not every task needs the largest model available. For narrow, high-volume work, a fine-tuned 3B model is often faster, cheaper, more consistent — and it can run on your own hardware.

16 July 2026Read

Agentic Commerce8 min

Designing a storefront an AI agent can actually shop

Commerce is moving from pages to agents. Most storefronts are built entirely for human eyes — and an agent cannot use any of it. What has to change.

25 June 2026Read

Vector Search7 min

Choosing a vector database: pgvector or a dedicated engine

Most teams reach for a specialist vector database far earlier than they need to. Here is the honest decision boundary, and what actually changes when you cross it.

5 June 2026Read

Laravel + AI8 min

Adding a RAG assistant to a Laravel app without rewriting it

You have a working Laravel application and you want an assistant that answers from its data. You do not need a new stack, a Python service, or a rewrite. Here is the pragmatic path.

20 May 2026Read

Evals7 min

Evals are the product

Anyone can ship an LLM feature that works in a demo. Shipping one you can change without fear needs something less exciting: a test suite for a system that is not deterministic.

2 May 2026Read