betterCode(), GenAI Summit · Mannheim · June 10, 2026
Who We Are
Enterprise Architect at Deutsche Telekom with a focus on AI. She established the use of AI for modernizing legacy systems at Telekom IT GmbH and is currently shaping a group-wide architecture framework for multi-agent systems.
Architecture & ODA
Product Owner at Deutsche Telekom IT GmbH in the field of artificial intelligence. With her team, she works on AI-supported modernization of critical enterprise applications and operates a chatbot platform for intelligent knowledge management.
Product & DeliveryAgenda
Problem Statement
With the Migration Plugin, we want to tackle the challenges of legacy modernization to increase the innovation cycle.
Legacy Systems
Deutsche Telekom Approach
Migration Pluginfor AI assistants
Claude Code
GitHub Copilot
Cursor
DeepEval
Spec Kit
Cucumber
Migration Plugin · Multi-Phase Workflow
Live Demo
From legacy artifact to validated target increment: reverse engineering, specification, forward engineering and evaluation in one structured loop.
Success Stories
“Your plugin developed a running feature that is now ready for testing.”
65,000 lines of legacy code were refactored into 4,400 lines. A new feature powered by a modern API merges data from two databases.
“We delivered 8 to 12 months of work in just one month.”
35+ repositories and 4 million lines of code were analyzed. The team delivered end-to-end modernization of a component across repositories.
“Without your plugin we would not have met our timeline for migration.”
Business logic was extracted from 1.7 million lines of code and new software was implemented for three business processes.
Success Factors
Human in/on the loop, AI coding assistants, lossless work and quality shifted left.
AI buddies, clear priorities, small business processes and sequenced technical/business transformation.
Models, context, retrieval, reuse and monitoring are managed deliberately as an operating loop.
Scope, model access, secrets, least privilege, human gates and auditability as default controls.
Strategic Principles
Enterprise Scale
MARA Guardrails
The Magenta AI-Centric Reference Architecture provides comprehensive patterns, guidelines and best practices for building enterprise-grade AI solutions at Deutsche Telekom.
Steer decisions and owner accountability.
Reverse engineering, architecting and forward engineering.
Score each app L0-L4, identify modernization gaps, prioritize waves and track coverage.
Capability-by-capability plan, funding, ownership and controlled cutover.
Governed transition from legacy complexity to AI-first solutions.
AI Adoption Metrics
Key question: Are we industrializing AI-driven modernization and making modernization AI-first by default?
| Level | Adoption scoring |
|---|---|
| L0 | Manual modernization only |
| L1 | AI-assisted discovery / reverse engineering |
| L2 | AI-assisted architecture & specification |
| L3 | AI-supported forward engineering |
| L4 | AI-assisted cutover |
Cost Control
Goal: lower model usage, integration effort and repeated context work
Reduce spend by separating planning from execution, reusing tools and context, and routing each task to the smallest sufficient model.
Cost control starts by making the main drivers visible, then managing them deliberately inside the workflow.
Cost Control
Four moves translate platform choices into lower run cost and less repeated context work.
Design principle: keep reasoning scarce, retrieval targeted, tool execution deterministic and integration reusable.
Scaling Patterns
Goal: prevent wasted runs, rework and uncontrolled usage
Reduce repeated analysis, wasted tokens, unnecessary modernization work.
Structured workflows reduce failed runs, rework and unnecessary model calls.
Persist plans, analyses and interview outputs so the same information is not recomputed.
Focus on valuable business logic; avoid modernizing dead or unused legacy code.
Clear specifications reduce ambiguity, retries and costly implementation corrections.
Compactly represent code and dependencies. Faster navigation, duplicate-analysis reduction and hidden relationship discovery.
Less recomputation • Smaller context windows • Fewer abandoned implementation paths
Use this area to prioritize which code and evidence the AI should spend tokens on.
Scaling Patterns
Goal: prevent wasted runs, rework and uncontrolled usage
Shift review effort into executable checks and targeted model judging.
Automated validation catches modernization errors earlier and reduces manual review.
Reserve expensive reasoning models for scoring rather than every generation step.
AI-assisted labelling and review lowers the volume of costly manual annotation.
Write tests before implementation. Use test suites as automated gates for AI/code changes.
Impact: fewer regressions, faster debugging and less rework.
Translate requirements into Given/When/Then scenarios and executable acceptance checks.
Impact: shared language, less handoff ambiguity and reduced manual review.
Cost Control
Make every modernization workflow cheaper by design: measure spend, route intelligently, reuse context, automate repeatable work, then package what proves reusable.
Track tokens, model choice, context size and workflow execution.
Use the smallest sufficient model; reserve reasoning for complex decisions.
Pull only relevant legacy facts; reuse stored outputs.
Automate repeated steps with tools and scripts.
Turn proven workflows into skills, plug-ins and marketplace assets.
Cost Control
Use the operating loop with explicit guardrails so teams reduce model spend without losing quality or reusable acceleration assets.
Baseline cost per workflow before changing prompts or models.
Route low-complexity tasks to cheaper models and deterministic tools.
Retrieve only the facts needed for the next decision.
Promote proven patterns into skills, plug-ins and reusable assets.
Outcome: lower model spend, shorter modernization cycles, consistent quality and reusable acceleration assets.
Security Requirements
Code migration only: no runtime operations, production data or system administration.
Repositories are scanned and cleaned before migration. Agents never access secrets or credentials.
Route interactions through approved model hubs with governed providers and safeguards.
Restrict agents to approved repositories, branches, files, tools and actions.
Engineers approve generated changes before merge, deployment or irreversible action.
Trace prompts, model calls, tool actions, code changes, approvals and versions.
Treat generated migration changes like production code with review, tests and rollback paths.
How to Start
Choose a valuable feature, not a whole application. Define success criteria and scope.
Load code, docs, APIs, UI traces and existing tests. Keep context targeted.
Use agents / coding assistants to extract flows, dependencies and business logic.
Create specs as guardrails, design target architecture, generate tasks and code.
SME review, tests, security checks and LLM judge. Package reusable skills.
Output: validated modernization backlog + first generated target increment + reusable workflow assets.
Wrap-up
Do not wait for the perfect market solution. Use current coding assistants and agents on a pilot.
Reverse engineering, specs, forward engineering, validation and human approval belong together.
Turn proven workflows into skills, plug-ins and guardrails to lower cost and improve quality.
Thank you
Time for questions.