Inside Deutsche Börse’s GenAI‑Powered Migration of 2,000+ Zeppelin Notebooks

The migration was driven by the upcoming retirement of Zeppelin within the Cloudera ecosystem. Cloudera documentation states that the Zeppelin notebook environment has been deprecated and is no longer supported in newer runtime versions, making long‑term reliance on it risky.

For Deutsche Börse’s StatistiX team, the challenge was scale: more than 2,000 notebooks contained analytical workflows and business logic accumulated over years of development. Rebuilding them manually would require a massive amount of engineering effort and would be difficult to coordinate across users and teams.

The Key Design Insight: Separate Structure From Logic

Rather than trying to automatically translate entire notebooks with a single AI prompt, the team designed a system that separates two fundamentally different tasks.

1. Structural conversion (deterministic)

Certain elements of a Zeppelin notebook are predictable and mechanical. These include:

• Converting Zeppelin paragraphs into Databricks notebook cells
• Translating interpreter syntax
• Reformatting metadata

Because these transformations are rule-based, they can be handled with deterministic code and automation rather than generative AI.

2. Logic reconstruction (AI-assisted)

The real complexity lies in the analytical logic inside notebooks. Business logic, data transformations, and analytical intent often require interpretation. Instead of attempting a risky automated rewrite, the app uses generative AI to guide users through rebuilding this logic.

Context-aware prompts are generated for Databricks Genie, helping users reconstruct the intent and functionality of each notebook step while keeping a human in the loop for validation.

This design makes the system more reliable: deterministic automation handles formatting and structure, while AI assists where interpretation is required.

Architecture of the Custom Databricks App

The migration tool was packaged as a Databricks App, meaning users could run the migration workflow directly within the Databricks environment instead of using an external tool.

At a high level, the architecture included:

• A user interface that guides users through notebook migration
• Automated processing that converts Zeppelin notebook structure
• A prompting system that generates contextual queries for AI assistance
• Human-in-the-loop validation for reconstructed notebook logic

This approach allows business users to participate in migrating their own notebooks while maintaining governance and oversight.

Step-by-Step Migration Workflow

The migration process typically followed four stages.

1. Notebook ingestion
Users export Zeppelin notebooks and upload them to the Databricks App.

2. Structural conversion
Automated rules convert the notebook shell—paragraph structure, interpreter syntax, and metadata—into a Databricks-compatible notebook format.

3. AI-assisted logic reconstruction
The system generates context-aware prompts to help rebuild the analytical logic using Databricks tools, guided by Databricks Genie.

4. Human validation and completion
Users review and finalize the notebook to ensure correctness, preserving domain knowledge and compliance requirements.

This hybrid workflow combines automation, generative AI, and human oversight.

How Much Time the Tool Saved

Before the tool was introduced, redeveloping a notebook typically required several hours of manual work.

With the new workflow, redevelopment takes about 15–20 minutes per notebook.

For a migration involving more than 2,000 notebooks, that implies:

• Roughly 500–667 hours of redevelopment work in the new system
• Potentially several thousand hours if the notebooks were rebuilt manually

The exact total hours saved were not publicly disclosed, but the reduction from hours to minutes per notebook represents a substantial productivity gain.

Why This Is a Practical GenAI Use Case

Many generative AI projects struggle to deliver measurable operational value. Deutsche Börse’s notebook migration stands out because it targets a specific, costly engineering bottleneck.

Several factors make the approach notable:

Clear scope – The project focused on a defined task: migrating legacy notebooks.

Hybrid automation – Deterministic rules handled predictable tasks, while AI assisted with interpretation-heavy work.

Human oversight – Analysts remained responsible for validating logic, which is critical in regulated financial environments.

This combination avoids the risks of fully automated code generation while still delivering significant productivity improvements.

Part of a Broader Cloud Transformation

The notebook migration also fits into Deutsche Börse’s wider modernization strategy. The company has been steadily shifting workloads to cloud platforms and expanding its data and analytics capabilities.

For example, Deutsche Börse re‑engineered critical systems such as its DAX infrastructure on Google Cloud, achieving results including 33% lower total cost of ownership, migration of more than 60 applications, and an 85% reduction in disaster‑recovery time for key SAP systems.

The group has also reported surpassing 50% of workloads running in the cloud, reflecting a broader push toward scalable, cloud‑based data infrastructure.

Within that transformation, migrating analytical assets like notebooks is essential. Without modernizing those assets, cloud platforms cannot deliver their full value.

The Takeaway

Deutsche Börse’s approach demonstrates a pragmatic model for applying generative AI in enterprise engineering:

• Automate deterministic tasks with traditional software
• Use generative AI for interpretation and reconstruction
• Keep humans in the loop where business logic matters

By combining these elements inside a custom Databricks App, the company turned a potentially massive manual migration project into a semi‑automated workflow—cutting redevelopment time to minutes per notebook while maintaining control over critical analytical logic.