Technology is the tool.
People are the goal.

Fibre rollout: technology and architecture advisory inside a live programme

Currently engaged, 2026 ongoing. I hold a senior programme role at a tier-1 fibre / Glasfaser operator in DACH, doing technology and architecture consulting with leadership responsibilities inside the programme across field-operations coordination and stakeholder alignment. Details sit under NDA, and a walkthrough is available after NDA.

Data warehouse without the cloud: AI on the client's own infrastructure

The problem. Most data warehouses either send AI queries out to a US-cloud API or skip AI altogether. For a mid-market DACH client with DSGVO constraints, competitive-data sensitivities and licence-cost pressure, neither option works. What we built instead was AI that runs on the warehouse itself, on the client's own infrastructure, with audit trails and no data leaving the building.

The result. Retires per-seat BI and CRM licences, keeps every AI query inside the building, and gives the business a documented exit path from its incumbent SaaS.

Under the hood: A read-only ki_middleware gives local AI safe, audited access to the warehouse. dlt pipelines ingest CRM, ERP and Excel-controlling data into a PostgreSQL DWH with thirty-one governed views. Dagster orchestrates, Metabase visualises, and a nightly sync feeds the open-source CRM. Containerised, ADR-documented, with an eight-chapter operator handbook and a DSGVO security policy. Open-source throughout.

The same principle, local AI with no data leaving the building, runs in the open in Throughline, a vendor-neutral memory store for nine AI assistants. PostgreSQL 16 with pgvector again, with no account, no telemetry and no cloud. Beta, 752 tests, MIT licence.

Head of Technology: mid-market DACH technology firm

I led technology strategy and operations through a phase of organisational transformation. The priorities were closing the strategy-delivery gap and modernising the tooling that was holding the business back, while the in-flight programs had to be kept stable at the same time. Specifics on team sizes inherited, targets and adoption metrics are available on request under NDA.

Managing Director, CRM & Data: mid-size German IT services firm

Managing-director mandate, two and a half years. Refocused the business toward CRM and larger B2B accounts, set up a CRM consulting practice, added offshore delivery capacity, and brought AI into the delivery process. Material EBITDA improvement over the period. Figures available on request under NDA. M&A screening and negotiations alongside.

Head of Data Solutions: European independent index provider

Led the Data Solutions organisation for an independent European index provider. Redesigned the data architecture in parallel with live operations, introduced ITIL v4 and Six Sigma (VoC, CtQ, KAIZEN), and set up a data governance framework fit for the regulatory environment.

CRM cloud migration: five regions in parallel, no sales-cycle outage

Program lead on the migration of local sales systems into a unified cloud CRM on Microsoft Dynamics 365 / Azure. Five main regions (US, Hong Kong, France, UK, Germany) ran parallel cutovers. Fifteen country sites in total. Sales processes reworked during the migration, sponsor coalition anchored in the regional heads. Delivered to plan across the five regions. No sales-cycle outage.

Investment Restrictions: three years on Bloomberg AIM, no portfolio incident

Three-year program migrating asset- and investment-management locations to Bloomberg AIM (Compliance Manager and Violations Manager). Responsible for test planning and governance across trading desks. Adoption protected through parallel-run periods and desk-level reinforcement. Three years, zero live-portfolio incidents.

Every tolerance-breach alert in the run pointed to a person on a desk, a trader who had to stop what they were doing and work with us. That's the adoption cost behind the zero-incident line, staffed for and reinforced week by week.

Solit Finance: from founding team to 110 staff in eleven years

I spent eleven years building and running Solit Finance, a DACH consultancy serving financial services and telecommunications. I led the operational business, acquired and ran client programmes, and handled M&A screening and negotiations alongside delivery. Headcount grew from the founding team to 110 people over the decade. Solit was the principal vehicle through which the 2010–2021 client programmes listed here were delivered. The mandates from 2022 onward are independent engagements.

Those eleven years shaped the lens through which every mandate since has been read. I ran the P&L, learned the craft of sponsor coalitions, and made the hard calls on when to invest and when to stop.

Next Generation IT: CRM and customer-data architecture across five countries

Principal consultant on the group's Next Generation IT initiative, part of the program-management leadership team. Transformation of the AppCom architecture (CRM & Customer Data Platform) across five countries, cross-country IT licence management, and the design and rollout of the NG IT architecture in Hungary and the Czech Republic. Infrastructure advisory to the regional CTIO/CTO.

Head of CRM West: an 80-person practice with P&L responsibility

Led an 80-person CRM consulting practice for Germany's west region. Principal consultant on CRM analyses and implementations, with business-development and P&L responsibility for the practice.

vibrasim: twenty pre-registered experiments, one clean negative result

For this project I deliberately picked a subject I have no training in, roughly four disciplines past anything I have ever learned. Inside a cube of 60 by 60 by 60 vibrations there is nothing but frequency, polarity, position and velocity. Local binding rules turn those into electrons, then atoms, then molecules joined by bridges that behave like synapses.

The real object of study is not the simulator. It is the question of which moves I reach for when none of my usual moves are available. For thirty years, every deadlock in client work had a shortcut through the domain. Here there is none, which is what makes the deadlocks clean.

On 31 May 2026 an autonomous run worked through twenty pre-registered experiments, each with a matched control and a recorded verdict. The result is negative and stands unchanged in the repository. The substrate delivers three of the four properties a selective memory needs, and the fourth under none of the levers tested. The acceptance bars were locked in advance and were not moved afterwards. Documented just as openly: the words vibration, electron and atom are names for abstract nodes here, not the physics they sound like.

The more durable half is the method itself. It has been split out into its own repository, where it runs as a pre-registered research pipeline for a single machine.

Souprise: answers over business data without a single invented figure

A question-answering system for invoices, orders and customer records that copies every value verbatim from the record rather than generating it. Totals and averages are computed in code, and the language model only writes the sentence around them. Where a record is unknown, the system declines to answer. Value accuracy of 1.000 across sixty test questions, with every acceptance bar locked before the tests ran. It runs entirely on your own machine, with no account and no cloud.

German LLM: a transformer built from scratch, every component readable

The problem. Most introductions to transformers use English, and most off-the-shelf models treat German as an afterthought. Neither helps anyone who wants to understand why attention behaves differently for a language with compound words, four cases, and verb-final subordinate clauses. This repo is the from-scratch build that makes every component legible.

A ground-up transformer for German, written to keep a gestalt understanding of what these systems actually do. Six tutorials walk through transformer mechanics, attention, feed-forward layers, German-specific tokenisation, the training loop, and data collection. A mini-dataset builder and training script sit underneath. Runs entirely offline.

EquiML: one dataset in, one report out

The problem. Fairness tooling for ML is scattered across libraries. One for metrics, another for constraints, a third for explainability. Every audit I ran turned into the same copy-paste glue code. EquiML is the three-command pipeline that replaces it. One dataset goes in, a report comes out, the accuracy-vs-fairness tradeoff made explicit.

A command-line tool wrapping fairlearn, SHAP, and scikit-learn into one pipeline: load, detect bias, train a fair model, compare results. HTML report, metrics as JSON, CLI and Python API. CI-tested, MIT-licensed.