— AI & Consulting

Artificial intelligence consulting for Munich – based near Ingolstadt

I build privacy-aware solutions for organizations in Munich and Bavaria. I am based near Ingolstadt and work on site in Munich or remotely for consulting, workshops, and implementation. Arrange an initial consultation to discuss your project.

— Three offers for your project

Private model server

Run models on your own infrastructure or in a location you control. I plan deployment, access boundaries, measurement, and a secure path into operation.

Data curation and small-model fine-tuning

I audit and curate your data, separate training from evaluation, and adapt small open models with traceable results when testing shows a measurable advantage.

Vertical integration

I connect models to your documents, tools, and workflows. Human approval gates, reliable checks, and clear fallback paths are part of the design from the start.

Private model server deployment → Data curation and small-model fine-tuning → Vertical artificial intelligence integration → Network Day-0 configuration →

— Where AI pays off

Where AI pays off

In Germany AI use is rising fast; missing skills and privacy concerns still limit depth. Globally, agents pay off when work is repetitive, document-heavy, and reviewable — Cursor, MCP, and RAG with verify loops, not chat demos.

— How this helps daily work

How this helps daily work

Network Day-0 / staged configuration

Template rebuilds and complex CLI jobs burn junior change windows across vendor fleets.

Fixed-price catalog jobs with AI-assisted console work and a human gate — Phase 1 list-price Cisco IOS XR and IOS XE.

Inbox and reply drafts

Hours in email, uneven tone, the same thread shapes every week.

An agent drafts replies in your voice for approval — faster, never sends on its own.

Calendar and invite triage

Invite pile-ups and manual weighing of rules and priorities.

Rules-based triage: accept/decline suggestions so the calendar stays usable.

Repetitive portal clicks

Training portals, forms, and internal UIs burn minutes on every pass.

Browser agents run the click paths; you only handle exceptions.

Scanned documents

Contracts and statements as image PDFs — not searchable, painful to extract.

OCR plus structured extraction: searchable text and less retyping.

Engineering with Cursor agents

Generic AI invents APIs and breaks production systems.

Rules, skills, and verify loops: repeatable agent work with measurable checks.

Internal APIs via MCP

Know-how is trapped in CLIs and GUIs; only specialists can move.

An MCP server over your APIs: natural-language questions, actions with verification.

RFQ compliance with RAG and agents

Hundreds of compliance questions in Excel; evidence spread across PDFs and decks; manual scoring takes days.

RAG over product docs plus MCP tools and a local LLM: evidence-backed verdicts with an audit trail; super-agent eval replaces copy-paste. Helps human compliance writers answer thousands of hard questions quickly and ground each answer in many different data sources.

Internal knowledge Q and A

Years of expertise sit in drives and mail threads; new hires ask the same questions every week.

RAG over internal docs with citations; specialists only review the hard edge cases.

Codebase migration with agents

Large migrations stall in planning; hand rewrites burn months of calendar time.

Cursor rules and verify loops move migrations faster while humans keep the review gate.

Security questionnaire drafts

Vendor security forms repeat policy text across dozens of portals every quarter.

First drafts grounded in policy docs; humans approve before anything is sent.

Telemetry-to-action workflows

Operational signals are spread across dashboards, command-line tools, and specialist teams.

Agents collect evidence, explain likely causes, and propose verified next actions while engineers retain the approval gate.

Reusable technical enablement

A successful workshop or laboratory session remains with the people who attended it.

The workflow becomes a reusable skill with prerequisites, checks, rollback steps, and clear handover material for the wider team.

— Selected reference cases

Selected reference cases

These examples show internal engineering and reference implementations. They do not claim paid customer engagements and deliberately omit confidential clients and systems.

Internal engineering

Evidence-grounded drafts for extensive questionnaires

Problem

Technical questions arrive in spreadsheets while reliable evidence is spread across many documents. Manual research is slow and difficult to review consistently.

Approach

Document retrieval gathers relevant sources, prepares a grounded draft, and holds every answer for human approval.

Outcome

The workflow produces reviewable drafts with sources and a clear handover to the responsible person. It does not replace technical or commercial judgement.

Reference implementation

Safely bounded tools for assistants

Problem

Specialist knowledge is split across several interfaces and command-line tools. A freely acting assistant would have excessive permissions and could carry mistakes forward unnoticed.

Approach

Narrow tools use validated inputs, least-privilege access, and checks against the actual result. Critical actions wait for your approval.

Outcome

The reference demonstrates a traceable workflow from request through verification, including a safe stop and a manual fallback path.

Internal engineering

Comparing small models with fixed test cases

Problem

One persuasive answer does not show whether a small model is dependable enough for a recurring workflow.

Approach

Separate training and evaluation cases compare the base model with an adapted version. Failure classes and known limitations are recorded explicitly.

Outcome

The result is a reproducible basis for choosing adaptation, deployment, or no fine-tuning at all — without an invented success metric.

Reference implementation

Structured extraction from scanned documents

Problem

Image-based documents are not searchable and are difficult to bring into a reviewable workflow.

Approach

Text recognition, structured extraction, and fixed plausibility checks prepare the content. Uncertain fields are marked for manual review.

Outcome

The reference produces searchable content and traceable review steps instead of passing uncertain results downstream automatically.

— Tools & stack

Tools & stack

Python · FastMCP · LlamaIndex · PyTorch · TRL/PEFT/QLoRA · Hugging Face · Cursor (Agent Mode) · MCP Protocol · Git/GitHub · SSH · Conda.

— How I work

How I work

Start small, protect sensitive data, verify early, and roll back often. I build systems that leave behind code, tests, documentation, and a repeatable path for the next engineer.

— Lab

Coming later

A small model on a Cisco router serial console

This is networking, not a chat demo: a small model types Cisco Internetwork Operating System commands on the serial console of a router. First-day bring-up and live diagnosis look the same on that screen — the device still shows a privileged prompt. They are not the same job. After configuration the model must stop. After a fault it must keep showing, pinging, and writing a report. One 1.5-billion-parameter low-rank model adapter could not hold both stop rules. The next shape is a sidecar: you name the job and paste this site's addresses each time; the matching specialist types on the console. The two model adapters already exist. Wiring them to a live router is parked for a later session.

Data curation and small-model fine-tuning → Request a consultation →

— FAQ

Frequently asked questions

What is an MCP server and when do I need one?
MCP (Model Context Protocol) is an open standard that lets LLM agents — Cursor, Claude, custom assistants — talk safely to external systems: databases, APIs, your own software. If colleagues run Cursor and need to reach internal infrastructure, a dedicated MCP server is usually the right architecture.
Do you work remote, or only in Bavaria?
I work AI engagements largely remote — that is the industry norm. Workshops and longer on-site engagements I prefer around Ingolstadt, Munich, Augsburg and Bavaria. Germany-wide by agreement.
What specific generative artificial intelligence experience do you have?
Full-time since 2024. Production Model Context Protocol servers over orchestration and planning interfaces, retrieval-grounded compliance assistants, local model evaluation and fine-tuning, and verified Cursor workflows. Before that, more than twenty years as a network architect.
What does AI consulting cost?
Hourly and day rates by arrangement — depending on scope, complexity, and travel. An initial scoping engagement (typically half a day) can often be flat-rated. Send a short description of what you have in mind and I will reply with an estimate.
When is fine-tuning not the first choice?
If current information, internal documents, or a few clear rules are enough, strong prompts and retrieval are usually faster and less expensive. Fine-tuning makes sense only when tests show a recurring quality gap and enough curated training data is available.

Request a Munich consultation

Briefly describe your project. I will reply with an initial assessment and help decide whether an on-site meeting in Munich or a remote session is the better next step.

Request a consultation →