// SYS_04 · integration · automation · internal tool
Six notices instead of three portals
How we turned three tender portals and a job portal into a short list assessed by AI — every hour, with notifications on Discord and by email.
// Challenge
The contracts exist, you just have to find them
A software company that wants to bid in public tenders has a simple problem: software contracts exist, but you have to find them. And there's nobody to sit in front of the portals every day.
That company was us. The notices we care about are scattered across three systems. BZP (the Polish Public Procurement Bulletin) lists national procedures under the Public Procurement Law (PZP). TED (the EU's Tenders Electronic Daily) collects those above the EU threshold. Baza Konkurencyjności (the EU-funded projects' procurement database) holds requests for quotation from EU-funded projects — contracts outside PZP that don't appear in the first two at all. To know what was happening on the market, we had to check three places and, in each one, scroll past notices for toner and power generator supplies.
We counted how much of it there is. In 72 hours the three sources produced 71 notices, and only 6 of them actually concerned our work. Six hits in three days is too few to spend time on three portals every day. And far too many to risk missing one.
A short list is only half the job, though. Each hit has to be opened, and the participation conditions and specification read, to know whether to bid at all — the notice title doesn't tell you that. On top of that come private IT jobs from a freelance job portal: there's no documentation there, but speed wins, and competitors' offers are public.
The constraints were hard: one person, zero budget, our own server. Plus three government interfaces with no contract and no guarantee that tomorrow they'll respond the same way as today. The more code we built on a foundation like that, the more we'd have to fix after every change on the government's side.
// Approach
One list instead of three portals
One list instead of three portals. Each source delivers data differently, so the radar brings all notices into one format. The devil was in the details. BZP writes procurement codes with a check digit, TED without it, so we compare just the eight digits. TED gives deadlines in a format that standard tools don't understand, so the radar has to parse it itself — otherwise active tenders silently drop off the list as closed. For a company, that's the difference between “there was nothing interesting” and “there was, but you didn't see it”.
Decision log
- ADR-01
Only what's worth reading reaches the inbox
The radar lets through notices with seven procurement codes (CPV) we picked for our work, then drops those whose title mentions supplies, purchases, hardware or servers. BZP filters on the server side, so in the test run it returned one notice, and it was a hit. TED matches codes too broadly: out of 70 notices, the local filter kept 5. Without it, every relevant notice would come with about fourteen to reject by hand.
- ADR-02
We rechecked what we had written off as impossible
The Baza Konkurencyjności interface has no documentation. At the first look, in July 2026, its search returned an error, so we treated it as closed and built a workaround: querying notices one by one, number after number — from a hundred to three thousand requests per run. In September we checked again. It turned out the search works without logging in if you pass one parameter, and without it, it simply crashes. Today a run is one request instead of thousands, and the whole workaround is gone from the code.
- ADR-03
Notifications that fit the way you work
A new match goes straight to Discord as a separate message with the contracting authority, codes, region and deadline — with a red border when fewer than five days are left. Email is a digest sent at 6:00 and 13:00, and only when there's something to send. The first email doesn't contain two hundred old items, because at start-up the radar marks existing notices as already sent.
- ADR-04
AI assesses, a person decides
An AI model reads every new notice — the full text, not just the title — and compares it with the company profile: what we do, what we don't, what references we have. It answers with a verdict of “worth it”, “to review” or “not for us”, in one sentence, with the main reason and the main obstacle. “Not for us” notices don't ping Discord. Under the rest there are two reactions to click: “yes” and “no”. Recent decisions go back to the model as calibration, so the assessment learns our choices.
- ADR-05
An agent reads the documentation for us
After approval, an AI agent finds the procedure documents itself, reads the specification, the description of the contract and the draft agreement — in PDFs, Word files and ZIP archives too — and sends back a verdict with three points: whether we meet the participation conditions, what the money is, and what the biggest risk is. The full report comes as an attachment. Because the agent opens addresses taken from other people's documents, it can't reach anything on our internal network.
- ADR-06
An offer written against the competition
For private jobs, we analysed 1,706 offers submitted under 385 software jobs. Most restate the client's requirements and claim years of experience; almost nobody asks questions, gives a schedule or states support terms after launch. After approval, the AI writes an offer that reads the offers already submitted for that job and goes one step further, and a third reaction sends it from the company account. An offer with an empty field — price, signature, link — won't be sent.
- ADR-07
With three government interfaces that can change without warning, the radar deliberately has little code — fewer places to fix after a change on the government's side
// figures
Read from production
- 71 → 6
- Notices from three portals → real matches in 72 hoursTrial measurement on 16 July 2026, without Baza Konkurencyjności; a one-off sample, not an average
- 47 → 10
- The same measurement after adding Baza KonkurencyjnościRepeated on 20 September 2026; 2 of the 10 matches came from Baza Konkurencyjności
- 70 → 5
- TED noise filtered out before sendingTED matches codes too broadly; each hit came with about 14 unneeded notices
- 3
- Tender sources in one listBZP (national), TED (above the EU threshold), Baza Konkurencyjności (requests from EU-funded projects outside the Public Procurement Law)
- < 5 days
- Red border when a deadline is closeA new match goes straight to Discord with the contracting authority, codes, region and deadline; an email digest at 6:00 and 13:00, only when there is something to send
Instead of three portals, we have one short list. In the measurement of 16 July 2026, the radar cut 71 notices from 72 hours down to 6 real matches. In September, with Baza Konkurencyjności already included, 10 of 47 notices remained, 2 of them from Baza — requests that don't appear in BZP or TED at all. These are one-off samples from live sources, not a monthly average.
The radar has been in production since 31 October 2026. In the first run, the AI assessed 5 new notices: 3 went to Discord as “to review”, and it rejected 2, giving the reason. We tested the whole chain live: from approval to a verdict based on the documentation took about 70 seconds. On the first day, three AI-written offers, prepared against competitors' offers, also went out from the job portal.
It's too early for the number of matches per month or measured time savings — the first day of production isn't a statistic, and we don't claim anything of the kind.
We already know one limitation: the code list is too narrow for Baza Konkurencyjności. Seven codes catch 7 open notices there, while the codes actually used in Baza would give 102 — the AI assessment lets us widen this filter safely.
The next two to three months are for collecting data. Then we'll decide on further development. If it turns out to be better to buy a ready-made subscription tool for about PLN 150 a month, we'll buy it instead of extending our own. First, the radar has to tell us whether it's worth it at all.
Technologies
- Node.js 24
- TypeScript
- MongoDB
- nodemailer
- Docker
- Discord webhook
Team
Konrad Barejko — End to end
// Related
- ServiceProcess automationProcess automation for companies of 30–300 people: n8n and custom code, ERP and banking integrations, a process audit with ROI valuation.
- ArticleAPI — what is it? REST API, webhooks and OpenAPI explained for businessAPI explained with NBP, KSeF and VIES as examples: REST API, webhooks, OpenAPI, API keys and integration security for business.
- ArticleBusiness process automation — examples and where to startBusiness process automation: how it differs from RPA and AI, Polish data, KSeF e-invoicing, our hands-off funnel, examples by department and the first step.
15 minutes. Zero selling.
Is your data spread across several systems? In 15 minutes we'll tell you whether it can be brought together into one stream.