About this project

I am Federico Giudici, a procurement professional working for an international organization in Geneva.

I started this project for two reasons:

• to learn, hands-on, what AI tools make possible for someone who does not write code;
• to test whether public procurement information can be made more usable and accessible to procurement professionals.

This is a personal project. LinkedIn

The site was built with Claude, Anthropic’s AI, based on my own research and my experience in procurement.

I acted as the product owner. I have no coding knowledge, but I have worked as a product owner in systems development. Claude wrote all of the code, the automated tests and the technical documentation. I did not write any code.

What I did by hand:

• set the objectives and the functionality the site should have;
• researched and chose the data sources;
• answered 49 rounds of decision questions — methods, wording, data, launch timing;
• created the accounts, paid for the services, and entered passwords and keys;
• set up the domain, published each change, and ran a few long jobs on my own computer;
• reviewed the pages and reported what was wrong.

Most of the build ran as scheduled Claude sessions. Each session read a written progress log, did one planned task, ran the tests, and recorded what it did and what came next.

• Started: 7 August 2026.
• Public launch: October 2026, after the monthly data refresh of 15 October — about ten weeks after the start.
• In that time: more than 230 changes to the code, about 56,000 lines of code, and nearly 2,000 automated tests.

Prices and statistics do not use AI. Unit prices, medians and ranges are plain arithmetic on published records, and every figure links to its source record. AI is used around them:

Live counts per source are on the Sources page.

Used. Costs in US dollars. “Development” is what was paid up to 2 October 2026; “Ongoing” is the expected running cost.

Data (free): procurement records published under the Open Contracting Data Standard by seven national procurement agencies, and the UN Operational Rates of Exchange for currency conversion.

Considered, not used.

Plus a Claude Pro subscription, which I also use for other work.

1. Re-check what was reported as done. An AI can report a task as finished when part of it is missing. Ask for the evidence, not the summary.
2. Check security on purpose. A dedicated security review found gaps that normal development had not.
3. Test with real data, in small batches first. A sampled measurement was wrong until the whole file was checked. One translation run failed on every item and still cost money; a test of 1,000 items found it before the full run.
4. Real usage beats estimates. The translation was estimated at 10–15 dollars and cost about 21.
5. Some knowledge of systems is necessary. AI often does not choose the most efficient way, and you need to recognise it. Running AI models on your own computer also needs at least basic PowerShell.
6. Cost needs explicit work, and iteration. Server usage, database size, heavy calculations and what can run on a local AI model all had to be worked on deliberately, often more than once, before reaching a better solution.
7. Plan major functions in advance. Translation was added late and first applied to the wrong text, so it had to be redone as one large pass.
8. AI sessions need a written hand-off. Each session starts with no memory. A progress log and one task per session kept the work on track.
9. Ask for simple reports. Short sentences, what happens first, and a clear list of what I had to do saved more time than any tool.