Trenify.org
Trenify Foundation project · a Polish AI model for classifying letters

Jerzyk sorts incoming letters

It recognises the type of a Polish letter, checks whether a deadline is running and rates how urgent the matter is. It helps with the first pass through the mail at a law firm, public office or accounting firm, so that matters with a deadline end up at the top of the pile.

It runs on the firm's or office's own computers, so letters are not sent anywhere. It can be connected to a document workflow system. Setting it up needs help from IT.

Version 1.0.0 · results from 20 October 2026 · author: Trenify Foundation · Apache 2.0 licence

← Local AI at Trenify Foundation
Example

See what Jerzyk picks out of a letter

  1. A letter arrivesFor example, a request to complete missing documents.
  2. Jerzyk points out what mattersDocument type · deadline · urgency
  3. A person checks the resultWhoever handles the matter decides what happens next.
Letter excerpt (Polish) Sygn. akt I C 412/26 … wzywa się pełnomocnika powoda do uzupełnienia braków formalnych pozwu … w terminie 7 dni od doręczenia wezwania pod rygorem zwrotu pozwu.
Model answer
Document typecourt summons99.8%
Deadline runningyes91.2%
Urgency (0–3)3, within 7 days92.8%
Due datedelivered 3 Nov → due 10 Nov 2026

This is a sample result from the documentation. You cannot upload your own letter on this page. The excerpt asks the claimant's lawyer to complete the claim within 7 days of delivery.

The percentage next to an answer is the model's confidence for this one letter, not the accuracy of the whole tool. A person should check low-confidence results.

The exact date comes from a separate deadline calculator that follows Polish civil and administrative procedure, including Saturdays and public holidays. The model itself does not compute dates.

Jerzyk is not legal advice and not the only safeguard for deadlines. It helps order the queue of letters; a person stays responsible for recording and tracking each deadline.

Scope

What Jerzyk checks

The set of questions depends on the type of correspondence. The model does not write replies or summaries. It always picks one of the given options and says how confident it is. You can also ask your own questions with your own list of answers.

CorrespondenceQuestions
Court lettersdocument type (e.g. judgment, decision, payment order, summons, hearing notice), deadline, urgency
Administrative lettersdocument type (e.g. decision, procedural decision, request to complete, notice), deadline, urgency
Accounting documentsdocument type (e.g. invoice, corrective invoice, bank statement, tax or social insurance office letter), cost category, deadline, urgency
Customer e-mailsdepartment (e.g. sales, complaints, payments), cancellation, upset customer, urgency

Urgency on a 0–3 scale: 0 – nothing to do, 1 – no deadline or more than 14 days, 2 – 8 to 14 days, 3 – within 7 days or immediate.

Results

Jerzyk 1.0 test results

  • 85.5%

    Mean accuracy on real documents. The test covered 145 letters: 89 court rulings from the SAOS database and 56 public notices from samorzad.gov.pl. Results are averaged first within each type of letter, then across types. For comparison, Qwen3 8B scored 83.0%, Bielik 11B v3 82.5% and basal-2.0-mini 78.1%. Results are on the published version of the test, in which personal names and addresses are anonymised.

  • 0.02 s

    Median time per letter. Jerzyk, Qwen3 8B and Bielik 11B ran on the same laptop, a Lenovo Legion Pro 7 with an NVIDIA GeForce RTX 5080 Laptop GPU (16 GB), and Bielik 11B needed 0.4 s. Jerzyk also runs on a CPU alone, without a graphics card, but more slowly. Speed then depends on the hardware.

  • 322M

    Parameters. The model files take about 650 MB.

Method and full results →
Limitations

What Jerzyk cannot do

  • Polish only, and it reads up to about 3,000 characters of a letter. It cannot read scans without a text layer.
  • In long administrative notices it finds the deadline in about 71% of cases. Bielik 11B reaches 79% here.
  • Its weakest result is urgency on real notices: 58.9%, on a sample of 56 documents.
  • We built the benchmark ourselves. Some test documents were written by Bielik 11B, which is also one of the compared models.
  • It is not legal advice and not the only safeguard for statutory deadlines. A person should check low-confidence answers.
Getting started

How to start

Jerzyk is a model, not a ready-made app. You run it in Python or build it into an existing document workflow. The model files download from Hugging Face on first use. Due dates come from the jerzyk.deadlines module.

pip install "jerzyk @ git+https://github.com/Trenify-Org/jerzyk"

from jerzyk import Jerzyk
j = Jerzyk("trenify/jerzyk-1.0")
j.ask("<letter text>", domain="court")

The model builds on Laya (Apache 2.0) and mmBERT (MIT). Details are in the README on GitHub.

Want to test Jerzyk on your own letters?

We are looking for Polish public offices, law firms and accounting offices to help us check how the model performs in everyday work. Write to us.

Contact us