BPMN-AI

A bachelor thesis project at IWB that turns written process descriptions into valid BPMN 2.0 models. The system runs locally, evaluates open-source language models and combines LLM reasoning with deterministic validation, XML generation and layout algorithms.

Build a local AI system that helps process owners and analysts create BPMN 2.0 models faster. The goal was also scientific: evaluate which open-source language models and prompt strategies work best for transforming process descriptions into structured process models.

BPMN-AI reduces the blank-canvas problem. A process owner can provide a description, the system creates a first valid draft, and the human can focus on review and improvement. The architecture avoids asking the LLM to generate everything. It uses language models for semantic understanding and deterministic code for what machines can do more reliably.

Why this project was needed

IWB already has many process models, but not every department has documented its processes yet. That matters because process models are the basis for automation, process analysis, consistency checks and finding inefficiencies.

Creating a process model is rarely a one-shot task. The process owner may forget a step, the analyst may need clarification, and gateways or alternative paths often require several review loops. BPMN-AI was built to shorten the first-draft phase without removing the human review step.

A JSON-first pipeline, not a black box

Built for confidential process data

The system was designed to run locally because process descriptions can contain confidential operational knowledge. Sending that content to a cloud model was not the target architecture.

A key part of the bachelor thesis was evaluating open-source language models and prompt strategies. The portfolio version explains the method and architecture at a high level, while detailed evaluation data and internal process examples remain private.

From prompt to BPMN, step by step

Specialist validators, not one judge

Why JSON first - feel the difference

The hardest part: a diagram humans can read

A technically valid BPMN model can still be visually useless. The layout stage was one of the most difficult parts of the project: lanes, tasks, events, gateways, documents and systems have to be positioned so the diagram is actually reviewable - no overlapping elements, lanes sized to their content, data objects placed next to the tasks they belong to, and flow crossings reduced as far as possible.

The output is not an image. It is an editable BPMN 2.0 file that renders in the frontend, can be refined by the user, opens in standard BPMN tooling and can be transformed for specific target modeling environments before publication.

From bachelor thesis to scientific paper

The project also became the basis for a scientific paper. The work has already been presented at the Society 5.0 Conference and is planned for publication on SpringerLink.

The paper covers the architecture, prompt strategies, model evaluation and governance impact of using local LLMs for BPMN generation. Publication link: coming soon.

What it delivers