Teaching an LLM how to Generate Axon Ivy Business Processes 🤖
If you’ve ever wondered what happens when you put two developers in a remote mountain hut 🏔️, add a dash of Swiss air 🇨🇭, and a pinch of AI ambition ✨, let us tell you: code magic ensues.
Armed with a vague idea and sophisticated JSON schemas that define Axon Ivy process models, we set out to answer a burning question: Can we teach an LLM to generate business processes for real-world applications? 💡
Our Mission: LLM-Driven Business Process Generation 🚀
This year, @Reguël Wermelinger and I dove deep into Axon Ivy’s core: designing business processes. Traditionally, crafting these processes required meticulous manual effort. But times are changing. With LLM-powered coding assistants like VSCode Copilot 💻✨, you can now describe what your process should achieve and watch it materialize almost like magic.
At the heart of Axon Ivy lies its JSON-based process models, which drive the Axon Ivy Designer, our visual tool for building workflows. But imagine taking it a step further: what if users could simply ask an LLM to generate a process based on their ideas?
👉 Enter the code camp! 🎒
The Axon Ivy MCP Server ⚙️
To make LLM-powered process generation more than just a cool idea 💭, we set out to build an MCP (Model Context Protocol) Server inside the Axon Ivy Engine.
An MCP Server can be used by coding assistants (MCP Clients) to gain domain-specific knowledge and invoke specialized tools 🛠️. This gives us much more control over the results of user prompts, especially for a domain as specific as Axon Ivy's business process models.
The Journey 🛤️
We started by creating a new repository 📂 and implementing the MCP Server infrastructure.
Once that foundation was laid, we explored its capabilities, developing simple tools and watching Copilot use them 👀.
We experimented with sampling requests, which allow an MCP Server to prompt an LLM via the MCP Client, so the server can leverage AI responses without needing direct access or credentials 🔑.
Then came the real challenge: developing a tool to generate Axon Ivy business processes that strictly follow our JSON schema 📜.
With some setup and prompt engineering, we quickly achieved usable results ✅. From there, we expanded our scope to also generate the necessary data classes, refining our tools, prompts, and schemas as we went.
And here’s what it looks like when it all comes together 👇
Naturally, we didn’t just code; we made sure everything was covered with tests 🧪, enabling rapid iteration with confidence.
Lessons Learned 📚
Collaborative Coding 🤝
Working closely pushed our ideas further. Regular discussions and brainstorming allowed us to rapidly experiment and develop new approaches.
AI Surprises 🎲
Getting an LLM to do exactly what you want is never straightforward. It takes a lot of prompt and tool flow experimentation. Sometimes, the LLM would ignore our tools entirely, proving that large language models are still a bit of a black box 🕳️.
Process Automation ⚡
We discovered real potential in using LLMs to automate repetitive modeling, freeing up humans for actual mountain hiking 🥾 (or at least more coffee breaks ☕).
What’s Next? 🔮
More LLM Tooling 🛠️
We’re just scratching the surface. Next, we could enable users to prompt for entire business flows, generating executable models with forms, variables, external system integrations, and even market connectors 🌐.
Product Integration 📦
The MCP Server isn’t yet integrated into the Axon Ivy Engine product. Making it an integrated feature would speed up development and allow us to gather valuable feedback for future improvements.
Continued Exploration 🧭
The AI space is evolving rapidly, with new products, technologies, and features emerging all the time. We’ll keep exploring different approaches and tracking what proves valuable, both internally and externally.
From the peaks of Switzerland 🏔️,
Reguel & Lukas ✍️
