Case · Automation and AI tender selection
19 UN portals in one morning list: tenders matched to one expert's résumé
Client: an external expert, an international consultant. The system is built around one specific résumé, so selection is pinpoint.
The task: send the expert a short list every morning of UN and development bank tenders matching their résumé, already checked and translated, without browsing the sites by hand.
UN agencies, development banks and bodies with international funding publish vacancies and consulting tenders on their own websites. There is no single window. Each has its own date format, its own way of showing a list, and its own protection against automated collection.
To find work for a specific specialist among them, a person would have to open two dozen sites every morning and read headlines behind which the substance is often not visible. Under "evaluation consultant" hides sometimes narrow gender expertise, sometimes public finance work, sometimes a completely different region. Half of what is found is useless: the deadline has passed or the country is wrong.
There are thousands of notices a month, and only a few suit this expert. Searching by hand means spending hours every day and still missing what matters.
We built a pipeline on our own code in which each step costs more than the one before, so anything unnecessary is cut off as early as possible.
Nineteen sources are processed one after another, from first to last. Each one goes through the whole path before the next one starts. If one site fails, the run doesn't crash: the other eighteen get to the end.
The same job is often posted on several platforms at once. The system recognizes the repeat, doesn't create it twice, and fills in the missing fields of the record from the second source.
Before the expensive step, every request goes through a geography verifier. If the country can't be determined, the request doesn't slip through to the AI blindly: either the location is clarified from the full text, or the request is dropped. Unwanted regions never reach the evaluation.
Artificial intelligence checks each request against the expert's full resume: fields, experience, languages, work region. The decision comes as a sentence with an explanation, so the reason for a rejection is visible.
What fits goes to a Telegram bot with a short summary in Russian and English. A working admin panel at a separate address shows everything: lists of selected and rejected items, a card for each request, and a live view of the run.
The hits are narrow and precise: the system is built around a single résumé, so it picks a handful out of thousands.
| Metric | Value |
|---|---|
| Sources monitored | 19 |
| Listings screened in 118 days | 18,699 |
| Filtered out by the prefilter before AI evaluation | 17,175 |
| Checked by artificial intelligence | 1,545 |
| Selected as relevant | 22 |
| Successful daily runs | 119 |
| Average run time | about 20 minutes |
| Hours of manual platform browsing | 0 |
| Working hours saved | 280 |
What gets selected are consulting contracts that fit the expert's profile: gender equality, social protection, program evaluation, public finance, the Central Asia region and neighboring countries. Uzbekistan, Kazakhstan, Tajikistan, Kyrgyzstan, Moldova.
Before: twenty-odd sites checked by hand every morning, hours per week, missed deadlines.
Now: a short morning list in a messenger, where every listing is already matched against the résumé and translated.
The system is running. Last run on 19.10.2026 at 06:00: 2,263 listings collected in 21 minutes. Database: 18,699 records, 19 sources; selection against the expert's résumé runs daily.
There are no financial figures here: the expert submits the tender bids personally, and we do not measure conversion into contracts.
The showcase is password-protected and not meant to be indexed, so it has no search rankings or mentions in AI answers, and it should not have any. All figures above relate to the operation of the system itself and were taken from the working database on 19.10.2026.
The same pipeline can be built for a different task in a few days. Three things change: the list of platforms, the selection rules, and the résumé or profile the AI checks against. The structure stays. We will review your task for free: see where your tenders are and estimate the build time. Reply within an hour, quote within 48 hours, and the price will not rise after approval.
The system monitors 19 sources: websites of UN agencies, development banks and bodies with international funding. Each has its own date format, its own way of showing a list and its own protection against automated collection. That is why 19 connectors were written in custom Python code, each source is handled in its own way, and layout breakages are fixed point by point.
In 118 days the system screened 18,699 applications. The prefilter discarded 17,175 of them before artificial intelligence evaluation, the AI checked 1,545, and 22 were found suitable for the expert's résumé. The funnel is so narrow because the system is built for one specific résumé and selects precisely, a handful out of thousands.
The nineteen sources run one after another, from first to last, and each completes the whole path before the next one starts. That way one site going down does not crash the run: the other eighteen finish. The sites also do not get simultaneous load, because the crawl is sequential and independent.
The same vacancy is often posted on several platforms at once. The system recognizes the duplicate, doesn't create it twice and fills in missing fields on the card from the second source. As a result, a vacancy published on three platforms reaches the expert's morning list once, with a complete card.
AI scoring costs more than the earlier steps, so anything unneeded is filtered out sooner. Every lead goes through a geography check. If the country cannot be determined, the lead is not sent to the AI blindly: the location is clarified from the full text, or the lead is dropped. That way, leads from unwanted regions never reach scoring at all.
Every request to the AI includes the expert’s full résumé: fields, experience, languages, work region. The AI compares the essence of the assignment against it, so selection is tailored to a specific person and doesn’t depend on keyword matches. This matters because “evaluation consultant” can hide gender expertise in one case and public finance in another.
The decision comes as a phrase with an explanation, so the reason for every rejection is visible. All rejections are saved in words, which shows exactly what cuts the selection. Rules are refined based on facts, and the admin panel has a separate rule editor for that. The operator understands why a request was rejected, with no guessing and no manual rechecking.
The Telegram bot is the expert's main channel. In the morning a short list of suitable consulting tenders arrives. Each one comes with a link, a brief summary in Russian and English, and a short explanation of why it matches the profile. Every tender is already checked against the CV and translated, so there is no need to open the portals yourself.
The working admin panel opens at shabesboy.com/tender behind a password. Inside are a feed of selected and rejected leads, a card for each with the AI's decision and translation, the selection rules editor and the "Logic" screen. There, the "Start" button shows a live run of the source: which leads pass each stage, which are filtered out and why.
A timer starts the system every morning at 06:00. The average run takes about 20 minutes, with 119 successful daily runs in total. The last run, on 19.10.2026 at 06:00, collected 2,263 listings in 21 minutes. Hours of manual platform browsing: zero, as the expert doesn't visit the sites.
Consulting contracts are selected to fit the expert's profile: gender equality, social protection, program evaluation, public finance. Geographically that is Central Asia and neighboring countries: Uzbekistan, Kazakhstan, Tajikistan, Kyrgyzstan, Moldova. Matches are narrow and precise, because the system is built around a single résumé and only passes through matches to it.
Yes, the same pipeline can be built for a different task in a few days. Three things change: the list of platforms, the selection rules, and the resume or profile the AI checks against. The structure stays. The studio reviews the task for free and estimates the build time: a reply within an hour, a quote within 48 hours, and the price doesn't rise after approval.
Tender applications are submitted by the expert himself, and the studio does not measure conversion into contracts. So the case shows only the system's own performance figures, taken from the working database on 19.10.2026. The showcase is password protected and not meant for indexing, so it has no search rankings or mentions in neural network answers.
Data is stored in SQLite with isolated access, so the project does not overlap with other systems on the same server. As of 19.10.2026 the database holds 18,699 records from 19 sources, and selection for the expert's résumé runs daily. The system is in production: the last run took place that same morning.
Bookkeeping app for a hair extension studio in Dubai
View case studyThe whole hosting was hacked, the site held up
View case study4 times more trainings and 25+ participants in each: 50+ pages on Tilda for the "Sila Soznaniya" center
View case study