Freelancer profile translated to English.

Description

Does your AI pilot work in the demo but isn't ready for reliable operation yet?

I identify cost, error, release, and compliance risks, and bring AI pilots into production in 2-6 weeks – with monitoring, clear responsibilities, and rollback.

I work with medium-sized and enterprise teams that already have a concrete process, PoC, or functioning workflow: AI Agents, n8n automations, RAG systems, Voice AI, or self-hosted models.

How I can support you:

1. Production-Readiness Audit

I check error cases, operating costs, approvals, monitoring, rollback, and GDPR, EU AI Act, and data residency gaps. You will receive a written assessment with prioritized risks and a concrete roadmap.

2. Production Build

I integrate the pilot into your existing system landscape – including APIs, access control, audit logs, error handling, documentation, runbook, and handover.

3. Specialized AI Systems

Voice and chat agents, PII redaction, RAG, multi-agent systems, and AI automation with n8n, Python, TypeScript, and APIs.

Proven results:

• German PII Redactor: Entity F1 from 0.375 to 0.917

• 100% valid JSON in the tested pipeline

• Self-hosted Voice AI: 0.3-second response time in an EU VPC

• 12 initiatives and €2.4 million portfolio transparently prioritized

• 19 years of enterprise IT experience

How I work:

Within 24 hours, you will receive a written scope with deliverables, milestones, acceptance criteria, timeline, and fixed price – or a clear no with justification. Implementation is done per approved milestone; each milestone can be stopped individually.

Technologies: Claude, Azure AI Foundry, n8n, Python, TypeScript, LangGraph, Qdrant, SAP BTP, HubSpot, Salesforce, and self-hosted LLMs.

Write me briefly about which process or pilot should become production-ready, what already works, and what is currently preventing go-live.

Industry field of expertise

Languages

German
Native or bilingual
English
Native or bilingual

Workplace preferences

Remote only

Primarily works remotely

Graphic Packaging International
Enterprise Architect | Head of AI
DIGITAL AND IT
September 2024 - Today (1 year and 11 months)
Berlin, Germany
Transformation portfolios are managed in Excel and reprioritized in quarterly offsites — budget follows the loudest stakeholder, not the greatest value contribution.

Solution: Composite Scoring Engine (4 weighted dimensions) + Capability Maturity Model (Process / People / Technology / Data) + AI Recommendation Layer with explainable reasoning + Live Portfolio Dashboards. Read/write integration with the EAM catalog (LeanIX, ServiceNow APM, Ardoq). Deployment on-prem or Private Cloud.

Result:
• 12 live initiatives scored across 4 dimensions, portfolio €2.4M
• Scoping: 4-6 weeks workshop → minutes (AI-suggested)
• Reprioritization: Quarterly offsite → upon any metadata change
• Transparent Allocation: 67% committed / 21% planned / 12% free

Stack: Composite Scoring, Capability Radar, LLM Recommendation Layer, LeanIX / ServiceNow APM / Ardoq APIs.
AI Automation IT Architecture Artificial Intelligence LLM Microsoft Azure
Graphic Packaging International
Solution Architect | AI Solutions
DIGITAL AND IT
February 2022 - September 2024 (2 years and 7 months)
Berlin, Germany
SAP production data flows into Dev/QA/Training. Deterministic masking tools (SAP TDMS, Delphix, Informatica) cover ~95% of structured columns — gaps remain in free-text NOTES, unclassified Z-tables, and OCR scans. A general LLM over 100% of the data is not economically viable.

Solution: Fine-tuned German PII Redactor (LoRA adapter on Qwen2.5-1.5B) as a drop-in layer AFTER deterministic masking. Routes free text and unclassified columns to the model, delivering structured JSON (redacted text, entity classes, risk level, review flags) for 12 German entity types. Runs entirely on customer hardware (150 MB adapter), GDPR-compliant by design.

Result:
• Entity F1: 0.375 → 0.917 (2.4x better than baseline)
• JSON parse validity: 81% → 100%
• Risk Level Exact Match: 0.50 → 0.94
• Training: ~90 min on a RunPod A40; Inference on a consumer GPU

Stack: Qwen2.5-1.5B-Instruct, LoRA / PEFT / TRL, Pydantic-validated JSON, HuggingFace Hub, Apache 2.0.
Artificial Intelligence LLM Agentic AI AI Automation Microsoft Azure
Graphic Packaging International
Information Technology System Administrator
DIGITAL AND IT
January 2012 - August 2020 (8 years and 7 months)
Berlin, Germany
Led the migration of 2,500 mailboxes across two Microsoft 365 tenants, covering 30 locations and 5 languages. Took the technical lead, built the project plan, delegated to external contractors, and coordinated rollout across time zones until every mailbox was moved.
Kubernetes Microsoft Azure Python Integration Engineer Software Development