Service 03
Data & AI
Operational data becomes decisions — pragmatic and GDPR-aware.
There is more in your operational data than a spreadsheet shows. I turn it into dashboards, reports and decisions — pragmatic and GDPR-aware.
Problem & approach
Most small companies sit on data nobody ever looks at: the POS system, the shop, bookkeeping, maybe machines or sensors — every system quietly collects, and in the end gut feeling decides anyway.
I don’t start with the technology but with the decision that should improve: which products carry the business? When does hiring pay off? Where are failures building up? That becomes a dashboard, an automated report, or an alert that speaks up before things get expensive — deliberately small, not a data project without an end.
I use AI where it carries weight, not where it shines: classifying text, enriching data, spotting patterns. My background for this comes from industrial data analysis — sensor data where mistakes were expensive.
Scope of work
- Dashboards & automated reports — live, as PDF or CSV
- Data plumbing — bringing shop, POS, sensors and spreadsheets together
- Anomaly detection & threshold alerts for day-to-day operations
- Pragmatic AI integrations — text, classification, assistance
- GDPR-aware delivery with EU hosting
How it works
- 01
The question
Which decision should get better? First the question, then the technology.
- 02
Data audit
What exists, what's missing, what can be connected — an honest inventory.
- 03
Build
Dashboard, report or model — deliberately small to start, quickly useful.
- 04
Handover & operations
Runs on your infrastructure or EU hosting — documented and maintainable.
AI that stays in the building
Most AI advice ends with the suggestion that you send your data to an American provider. There is another way: models run on your machine or on EU infrastructure — your calculations, your customer records and your recipes never leave the building. This is not a statement of intent. This site loads not a single third-party resource, sets no tracking and needs no cookie banner. I practise what I recommend.
-
Own hardware
35 billion parameters, in my own house
Qwen3.6-35B-A3B runs locally via llama.cpp — on my own machine, not on rented compute. It is the same machine this site is built on. No prompt, no document and no calculation goes to an outside provider along the way.
-
Product
One seam, two worlds
Kardia's AI assistant talks through an OpenAI-compatible seam. The same inference runs locally (llama.cpp) or EU-serverless — switching is a configuration change, not a rebuild.
-
Production
17,000 products, categorised locally
At MegaWahl a local vision model with constrained decoding decides which category a product belongs to — more than 17,000 live products, and every writeback only after human approval.
Selected work
Reference work
Tools & stack
- Python
- pandas
- PostgreSQL / TimescaleDB
- FastAPI
- scikit-learn
- LLM integrations (EU)
Sounds like your project?
Tell me briefly what it’s about — I usually reply within one working day.
Get in touch →