← All work

JT Digital · our own sales system · since March 2026

Kataster

The sales engine JT Digital built for itself reads the websites of 59,551 Slovenian companies on servers we run and drafts first emails that a person approves.

A JT Digital product

59,551

Slovenian companies in the knowledge base, each website read by our crawler

~103,000

Lines of TypeScript, about 48,000 of them tests

478

Finished emails from the first version, in 16 sessions run by two people

Kataster's six steps in two rows. A knowledge base of 59,551 Slovenian companies, a crawler that reads each website, a profile written by a language model, sorting and scoring against each offer, a first email with its facts checked, and a person who approves it. A built-in CRM for contacts, deals, tasks and calls sits under them.
Illustration The path every company takes through Kataster. Nothing is sent before a person approves it.

One app, seven stages

Kataster is JT Digital's own sales system. It keeps a knowledge base of 59,551 Slovenian companies and works through it in seven stages: discover, crawl, profile, classify, evaluate, draft and grade. A person reviews every drafted email.

It is one web app in TypeScript on Fastify and Postgres, rendered on the server with fluent-html and HTMX. It has about 103,000 lines of code, about 48,000 of them tests.

The crawler, on infrastructure we control

The crawler is a separate worker on servers we run in Germany, with no access to the database. It takes a company's domain from a message queue, opens the home page in a real browser, fetches the other pages directly and leaves the text in file storage. It asks a site for one page at a time, two seconds apart, and obeys robots.txt. From every site it extracts contact addresses, links, the technology the site runs on and its language.

Profiles and classification by a language model we host

Five of the seven stages use a language model. All five run on Qwen3 14B, an open-weight model, on infrastructure we control, and no company's website text goes to an outside AI provider.

The model writes a profile of each company and sorts it along ten dimensions, among them Slovenia's 12 statistical regions, 10 kinds of organisation and 7 sales channels. Every fact in a profile must share a number, a name or a word with the company's website, or it is dropped. The stages share four requests at a time and queue behind other JT Digital apps whose users are waiting.

Funnels and email

A funnel is one offer, with its own customer criteria, language, sending hours and daily caps. Every company is scored separately for every active funnel.

For a good match the model drafts an email, code checks it, the model grades it, and a person approves, edits or rejects it. Approved emails leave every five minutes through the funnel's own SMTP mailbox. Replies and bounces are read over IMAP every ten minutes, opens and clicks are tracked, and one suppression list covers every funnel.

The first version sent real email, 478 finished emails in 16 sessions between March and May 2026. The rebuilt sending path is finished and tested.

Email genomes

An email can be assembled from a genome, which holds one choice for each dimension of the email. A genome earns a fitness score from its emails' opens, clicks, replies and won deals, and loses score on bounces and unsubscribes. A genetic algorithm breeds the next generation of 40 genomes from the best ones when an operator starts a round in the admin. The first version drafted its emails this way, and the rebuilt drafting stage does not feed the genomes yet.

A CRM and an operator's dashboard

The same app is a CRM for contacts, companies, deals, tasks and calls. Companies can be filtered on all ten classification dimensions, saved as segments and exported as CSV. A Pipeline screen shows every stage with the reason it is waiting, and every paid service has an hourly and a daily spending ceiling.

Want something like this?

Tell us what you need and we build a working prototype first, so you see it before you pay for it.