From one sentence to a finished PDF — a small decision model instead of an LLM

Problem

Filling in forms from a DOCX template — for example an overtime work request — is repetitive and error-prone when it comes to dates and settlement options. A full generative LLM is expensive and slow for this, can make things up, and usually means sending the data to the cloud.

What I'm building

• A Polish sentence is broken down into closed multiple-choice questions: work mode, settlement method, year, month, day, hours.
• The questions are answered in parallel by a small local decision model, Kev 0.8B (open source, Apache-2.0). It only picks answers from an allowed set and returns their probability.
• When confidence falls below a threshold, the system asks the user instead of guessing.
• Finally, domain validation and a PDF generated from the DOCX template.

Status and outcome

This is work in progress: the API is under construction and there is no UI yet. The design goals: everything runs locally, no generative LLM, and answers come only from the allowed set. I'll publish response-time and accuracy measurements once the project is finished.