Mahdi Salmanzade

Writing by Mahdi Salmanzade

Field notes on the decisions behind applied AI, drawn together here from the CLRT Insights archive.

• 7 July 2026 · Building

  Claude Code Artifacts Will Replace Half Your Documentation. The Other Half Is the Whole Problem.

  Claude Code Artifacts turn documentation into a byproduct of the work instead of a chore that happens after it. The generation is the commodity. Deciding what deserves to be certified as true, and making a document an engineer or an auditor can actually trust, is the part that does not ship in the feature.

• 22 June 2026 · The Market

  Claude Tag: The Friction You Hated Was the Governance You Had

  Anthropic shipped Claude Tag in June 2026, and the market read it as AI finally getting easy: tag a model like a coworker, no tab required. The easy part is the install. The hard part is deciding what an agent may touch in the one place where conversation turns into action, and proving afterward that it behaved.

• 15 June 2026 · Building

  Fifty Workflows Is the Wrong Unit of Value

  The list of agent workflows is free, and Opus 4.8 makes it freer. What a list cannot give you is the judgment about where to point an agent and the engineering to make a business trust it in production.

• 6 June 2026 · Building

  Loop Engineering: Why a Self-Improving Quant System Is the Hardest Thing to Trust

  The model that writes a trading signal is now a commodity. The scarce thing is the engineering and judgment that decide which signals are ever allowed to touch real money, and almost nobody builds that part well.

• 27 May 2026 · Building

  Organizational Memory Is Not a Folder You Point AI At

  Every company wants AI that remembers its context instead of starting from zero. The trap is believing that memory is a storage problem. It is a truth problem, and trustworthy organizational memory is one of the hardest systems to build.

• 16 May 2026 · Building

  Autonomous Loops Are the Operating Layer. Trusting One Is the Hard Part.

  An always-on machine running AI loops is trivial to start and unexpectedly hard to trust. The model is the commodity; the scarce skill is knowing where to point an autonomous loop and engineering it to run safely once you are no longer in the room.

• 28 April 2026 · Building

  Hippo: The Go LLM Client I Built Because AI Apps Need Memory, Budgets, and Tools

  A pure-Go LLM client Mahdi built as proof of a single idea most teams miss: the model is the commodity, and the operating layer around it, memory, budgets, routing, tools, and privacy, is where the real engineering lives.

• 7 April 2026 · The Market

  MCP-Dubai: The Open Source Bridge Between AI Agents and the UAE

  The model is not what breaks when a question turns local. The data underneath it is. MCP-Dubai is proof of how differently that problem has to be solved, and why grounding an agent in the UAE is the hard part almost nobody is solving at this depth.

• 24 March 2026 · The Market

  Odysseus Has Landed: PewDiePie's Self-Hosted AI Workspace and the Local AI Shift

  A famous creator shipped a self-hosted AI workspace, and the crowd read it as permission to run their own. The real lesson is quieter. Owning your AI is an engineering problem, and the distance between wanting sovereignty and surviving it is where most teams break.

• 3 March 2026 · The Market

  Access Is Free. Judgment Is Not. The Real Cost of AI Is Pointing It at the Wrong Work

  The price of intelligence has collapsed to near zero. The expensive mistake is no longer the subscription. It is aiming a capable model at work that does not matter, and trusting an output that was never built to be trusted.

• 10 February 2026 · The Market

  The Stack Is Free. Knowing Where to Point It Is Not.

  Tooling and access have collapsed to nearly free for everyone. Which is exactly why assembling your own stack is no longer an advantage, and why most businesses are optimising the one thing that stopped mattering.