Denys Shkola
EmailCV (PDF)

Case study 07AI engineering · Quant

AI-assisted legacy modernization at Erste Digital

An AI pipeline that turns legacy Natural/Adabas banking code into as-is docs and review-ready Java/OSB integration code: 5 use cases, 10+ connectors delivered.

Employer work, confidential
Status
Employer work · confidential
Role
Software Engineer
Period
Ongoing · figures as of Aug 2026
Stack
Natural / Adabas (mainframe) · Java · Oracle Service Bus (OSB) · GitHub Copilot CLI · Agents and reusable skills · MCP · Bash orchestration · CI/CD · Static analysis

The problem

The legacy banking code I work with is Natural/Adabas on the mainframe. Building a Java/OSB integration connector for it starts with working out what the old code does: which programs call which, what the screens show, which business rules apply. Before the pipeline, a typical use case took more than three monthly release cycles. In a regulated bank, generated code needs evidence before anyone merges it.

What I built

I lead an internal AI accelerator team at Erste Digital, Erste Group’s technology subsidiary. Our AI modernization pipeline for this code produces as-is documentation and Java/OSB integration connectors, delivered as a pull request with complete, usable code for human review. I orchestrate its tooling in CI/CD: GitHub Copilot CLI, agents and reusable skills, MCP, Bash and a deterministic code transpiler.

How it works

Schematic of the pipeline in four steps: legacy mainframe code, AI-assisted analysis, quality checks with human review, and documentation plus modern integration code delivered as pull requests.
Fig. 1The modernization pipeline at a high level. Schematic.

AI agents read the legacy Natural/Adabas code, explain what it does and draft the documentation and the modern Java/OSB integration code. Automated checks and a human reviewer sit between every generated change and the main branch, so nothing merges on the model’s word.

Results

Figure
5complete use cases delivered with the pipeline
Figure
10+Java/OSB integration connectors delivered
Figure
>80%shorter typical use-case cycle time: months to under one work week (n = 5)

As of August 2026 (my own figures):

  • 5 complete use cases and 10+ connectors delivered with the pipeline.
  • Typical use-case cycle time down by over 80%: from more than three monthly release cycles to under one work week and a single release (n = 5).

What I learned

  • The gates are part of the product. In a bank, generated code is only useful if people trust it, so the checks and human review sit inside the pipeline.
  • Document before generating. With the screens and business rules written down first, code and documentation can be checked against each other.
  • Name the metric precisely. “Months to under a week” is cycle time; effort saved is a different claim, and I don’t make it.

Notes