The task
A supplier who bids on public tenders has to watch the state procurement portal every day.
The portal publishes about 40,000 procedures a year.
Most of them do not fit the company.
The documents are in Serbian, and the technical specification is often a long PDF file.
A tender manager can spend hours on search and still miss the right contract.
What the client gets
Tender matching
- The client describes the interest in a search profile, a template.
A wizard builds the template from a ready industry package in a few steps. - Matching has two stages.
First a filter: CPV code, region, budget, submission deadline and procedure type.
Then keyword rules: "any of" groups, "all of" conditions and stop words, with Serbian word forms. - All templates can be switched on, switched off or deleted in one place, the "My account" page.
Delivery
- Telegram: a tender card with action buttons comes right after publication.
- Email: notifications to the inbox.
- Web cabinet: a feed of all matched tenders, sorted by date, template or deadline, with an "all or current only" switch.
- If a channel breaks, for example the client blocks the bot, the service reports it through another channel and shows a banner in the cabinet.
Work with a tender
- A status funnel: Studying, Participating, Submitted, Won.
- A daily morning summary with the funnel counters and the five nearest deadlines.
- Deadline reminders.
A notification when the award decision is published. - A buyer page with analytics of that buyer's procurements, and an export of the client's own tenders.
AI on Claude models
- Translation of the tender card into the client's language.
- Analysis of the technical documentation on button: requirements, risks and deadlines.
The depth depends on the plan. - Also on button: a bid or no bid verdict, a checklist of documents for the bid and a draft of the bid.
All three work within the plan limit. - Every model call is checked against the plan limit before it starts and recorded after it ends.
When the limit is reached, the client sees it at once and does not get a silent refusal.
Public part
- An open tender catalog by industry, 43 CPV sections, and by region, 30 regions.
- A sitemap and markup for search engines and AI assistants: JSON-LD and llms.txt.
Architecture
Five decisions that shaped the system
1. One shared bot for all companies
An early plan was a separate bot for each subscriber.
With many companies, that means many tokens, many webhooks and many points of failure.
So there is one bot, and each company has its own profiles in the database.
A routing worker compares every new tender with all profiles.
2. The database itself keeps client data apart
Every client table has PostgreSQL row-level security rules.
The bot and the API work under a restricted database role.
This role is not a superuser and cannot bypass the rules.
The data of one client stays closed to another client even if the code has a bug.
3. Every stage is a separate service
Intake, normalization, matching, the delivery queue, AI and sending run as separate services, 19 in total.
A failure in one stage does not stop the others.
The delivery queue protects against repeated sending, so there are no duplicates and no losses.
The service does not send to a channel that is known to be broken.
4. AI on request, with a limit check on every call
The client presses a button, and AI reads the specification files.
The answer lists requirements, risks and deadlines, and its depth depends on the plan.
The same buttons give a bid or no bid verdict, a document checklist and a bid draft.
Each call passes a limit check before it starts and goes to the usage records after it ends.
AI cost stays under control, and the client always knows the reason for a refusal.
5. Care for the data source
Requests to the state portal go with a rate limit.
When the portal is overloaded, the intake slows down on its own.
From the first phase, the data model also has a country registry, so a new market needs a new source, not a new product.
How it was built
I am the architect and the owner of the product.
I set the goals, the limits and the budget, give tasks to people and AI agents and accept the result.
A coordinator agent checks that a task has real value in the data, then writes a short brief.
A worker agent writes the code and the tests in an isolated branch and opens a draft pull request.
A second agent reviews the change.
The coordinator runs the quality gate again and checks the live site after release.
Every change passes the linter, the formatter and the full test suite with a coverage threshold.
Every new test must fail when its fix is removed.
High risk changes go only through me: database migrations, deployment, sign in and billing.
751 pull requests were merged this way between 2 May and 20 October 2026.
The service runs on a VPS in Europe, deploys from the main branch with one command and is checked after every deploy.
Three incidents that changed the rules
- A report said the gate passed, but the gate ran before the last commit.
Now the gate runs after the final commit, and the coordinator runs it again. - A pull request was merged before anyone reviewed it.
Now every pull request is a draft, and a hook blocks the rest. - Six planned tasks were not real problems.
Now every task starts with a check of the real data.
Public code: github.com/TnSrb1/nabrs
Agent workflow: github.com/TnSrb1/agent-ops
Status
The platform works on real data in open beta, free during the beta.
By 20 October 2026 it has processed 60,500+ tenders and delivered 28,700+ notifications.
The code base has 1,329 commits since May 2026.
Plans: Free, Pro at 29 euro per month, Business at 79 euro per month, and Enterprise on request.
Stripe payments are built in the code: subscription checkout, webhooks, renewal and expiry emails.
They will go live when the company details are ready.
In development
- Semantic matching on pgvector.
The pgvector 0.8.2 extension, the vector fields for tenders and profiles and the similarity index already work in production.
The matching code has a third stage for it.
The next step is to compute the vectors and switch the stage on.
Then the service will find tenders that are close in meaning even without the same keywords. - AI email digests.
An email with the tenders of a period and a short AI summary.
The email channel and the daily summary are already there.
Watch, match, read, deliver.
The same pattern fits any stream of documents that a team must not miss.
