TENDER-IT: search for UN IT tenders

In 18 days TENDER-IT processed 9,019 listings from 16 international organization platforms and selected 178 orders suitable for the SEO7 studio. All selected orders arrived in the messenger with a ready Russian translation, and not a single hour went into manually browsing sites. On average the system found about 10 orders a day.

TENDER-IT collects tenders from 16 platforms: websites of UN agencies, development banks and European Union bodies. There is no single window: each platform has its own date format, its own way of showing the list and its own protection against automated collection. That is why the system has a separate, independent automation branch for each platform.

A person would have to open sixteen sites every morning and read headlines where the title doesn't match the substance. “Data management” can hide a job for a corporate data warehouse engineer, and a dry wording can conceal an order for a web portal. Half of what's found is useless because it has expired, and digital projects make up only a few percent of thousands of listings.

The TENDER-IT pipeline has five steps: collection, cleaning, loading descriptions, selection by artificial intelligence and translation. Each step costs more than the previous one, so anything unnecessary is dropped as early as possible. The cycle runs at night on a schedule, and in the morning the person gets a ready list of suitable orders without browsing the platforms by hand.

In the cleaning step, the system discards expired and clearly off-topic listings by title, before downloading full descriptions. This way, out of 9,019 listings, 5,927 were filtered out without a single paid operation. The full task text is downloaded only for the rest, and artificial intelligence ended up reviewing 2,632 listings.

To load full descriptions into TENDER-IT, four strategies are used for different platforms: direct download, anti-bot bypass, the internal interface of a recruiting platform, and a real browser in the cloud. The cloud browser is needed for sites that render the list with scripts and are protected. This way the system handles even the most difficult platforms to collect from.

The AI checks each task against the company profile: what the studio can and cannot do, which languages it works in and where it is ready to work remotely. The profile is passed with every request, so selection is based on the substance of the task. The decision comes as a sentence with an explanation, and the reason for a rejection is visible at once, with no guessing.

Full Russian translation is the most expensive operation in the TENDER-IT pipeline. So it runs last and only for selected orders. As a result, paid translation covers 178 orders instead of 9,019 listings. The user gets every matching order already translated, with an explanation of why it fits.

The logic is built with visual n8n automations: twenty-two branches grouped into five stages. Opening the diagram shows which block is running now and where a failure happened, without reading technical logs. Changes are made without a programmer, and on the «Logic» page the active branch is highlighted along with the result of the last run.

More than half of the matches came from the European Union tender portal: 104 of 178 selected orders. That makes sense for a company registered in Spain, but before the system launched this conclusion was not obvious. The other suitable orders came from other international organization platforms, out of the sixteen connected to the system.

In the very first weeks, the system found support for a ticketing system in Spain, the digital transformation of a Greek municipality, a web portal for a Montenegrin state agency, digitalization packages for Ukrainian entrepreneurs, and the development of a web module within a government platform. These are digital projects that are easy to lose among thousands of listings in a manual search.

The TENDER-IT site is split in two. The public showcase with live metrics is open to everyone at tender.seo7.es. The working part is password-protected: it holds the lists of selected and rejected items, a card for each order with the AI's decision and a Russian translation, a selection rules editor and a live pipeline diagram. A new word in the rules starts working the next morning.

Each platform's branch is independent, so one site going down doesn't stop the other fifteen. Six platforms start at five-minute intervals so the sites aren't loaded all at once. A repeat crawl creates no duplicates; only the date the listing was last seen is updated. The log records every run, and 619 have been completed so far.

As of 19.10.2026 the system is stopped by the owner's decision. The data, site and settings are preserved: 9,019 records and 178 selected orders with translations, and restarting takes a minute. There are no financial figures, since no bids were submitted for the tenders found. The showcase was not promoted, so it has no search positions either.