A custom AI tool is software built around the work of one company: its documents, its prices and the steps its people repeat every day, with a language model in the part that used to need someone to read, write or sort. A company needs one when the products on the market cover most of the job but not the part that makes it different, and when the people who do that part by hand can say what a good result is.
What I build
- Tools around a quote or a document. Cotizera is a mobile quoting product for small businesses in Latin America: branded PDFs, delivery by WhatsApp and a pipeline from quote to won. Contratua is a private workspace where B2B teams keep contracts, versions and confirmations in one traceable flow.
- Tools that read the web or publish to it. Crawlio is a macOS crawler with a built-in MCP server that clones sites to the local disk. notion-to-site is an open-source package that keeps the writing in Notion and the site in code.
- A module inside a product that already exists. For Talisman, an accounting platform from Texas, I built a project-management module end to end as a contractor: the daily feed, the board, the engine that schedules recurring work and the chat of its assistant.
- Tools for people and for agents at once. LeanManufacturing.mx gives a plant manager a maturity self-assessment and an OEE calculator, and gives an AI assistant a public MCP server over the same material.
When a custom tool is the right answer
Building software has become cheap, and that is the risk: when making a tool is faster than understanding the process, a company collects small tools and loses clarity. So the first question is whether the tool should exist. The signs that it should:
- The work repeats, and its steps can be listed: collect the requirements, write the document, send it, follow up, update the record.
- The cost is in dropped steps more than in slow ones: a follow-up nobody sent, a pipeline updated late.
- The tool will have one owner and a defined connection to the systems you already run.
- Someone can say what a correct output is, so the tool can be tested.
When the signs are missing, the useful work is to understand the system you already have.
How the work runs
- The path, written down. With the people who do the job, I write the path the tool will carry: what comes in, what is looked up, what goes out and who approves it.
- A specification for every feature. Each feature starts as a chain of documents: the requirement, a handoff that can be executed and the prompt that runs it.
- Boundaries in the repository. I build with several AI coding sessions in parallel, each in its own worktree under a written contract of what it may change. A guard before every commit rejects changes outside that area.
- Assembled apart, then merged whole. The architecture is put together and validated away from the busiest code of the product, and enters the codebase as one piece.
- Tests as the gate. Tests decide a deploy. The part a model does gets regression tests for each flow and its edge cases, and a record of tool calls and generated files that a person can review.
What you receive
The tool running under your accounts, its source code in your repository, the specification behind each feature, the tests and a short guide to deploy and change it. Where a model is involved, also the prompts and the cases that test them.
What this does not include
I do not put a chatbot on top of a product and call it an AI tool: the model goes into the repeated work around the task. A tool does not replace the person who holds the judgment. In a quoting tool, the seller keeps the price and the relationship. If a product you can buy already does the job, I say so.
Where the evidence is
The products named here are on my work page, each with its address. The case studies linked below cover Cotizera, notion-to-site, LeanManufacturing.mx and the Talisman module, which shipped in three months. Airtable MCP and MetaMCP, shown with this page, are open-source work of mine, with their limits stated next to them.