Writing

Twelve months ago I was skeptical. Here's how my day-to-day workflow shifted, and what I still refuse to hand over.

Enterprise AI gets the headlines, but a clinic or a law firm on WhatsApp taught me more about reliability than any big project.

Traces, not logs. How I instrument agent runs so that "why did it do that?" takes minutes instead of an afternoon.

Realtime speech models changed what's possible on a phone line. Gulf dialect, code-switching, and latency are where it gets interesting.

Multi-agent diagrams look great on slides. Most of the time a single agent with good tools wins. Here's when it doesn't.

Prompt injection isn't a bug you patch, it's a property of the system. How I scope tools and permissions so it can't do real damage.

Not every request needs a frontier model. A practical look at routing, caching, and measuring cost per resolved task.

I stopped treating evals as a research exercise and started treating them like unit tests. It changed how fast I can ship.

Dialects, code-switching, RTL quirks, and tokenizer costs: what I've learned building assistants for Arabic-speaking users.

MCP everywhere, coding agents in every editor, and models that can work for hours. A look back, and a few predictions I'll probably regret.

GPT-5.1, Gemini 3, and Claude Opus 4.5 all landed in November. Benchmarks won't tell you which one fits your work. Your own evals will.

October was a busy month for agent tooling. What I tried, what stuck, and why I still reach for code over canvases.

Models can now work on a task for hours. The hard part has moved to the harness: loops, verification, and knowing when to stop.

A frontier model and serious open-weight models arrived days apart. For teams with data-residency rules, that changes the conversation.

Prompt engineering was never really about clever wording. It's about deciding what the model gets to see, and that has a name now.

Million-token context windows didn't kill RAG. Here's the playbook I use for retrieval quality, evals, and fallbacks on enterprise data.

OpenAI adopting MCP was the moment it stopped being one vendor's idea. What the protocol gets right, and where it still hurts.

GPT-4.1, o3, and Llama 4 all shipped in April. If swapping a model scares your team, your delivery pipeline is the problem.