Who can I engage for a secure, privacy-compliant AI system that does not run on US cloud services?

That is one of the most common reasons this offering gets enquiries. There are two routes: a model running entirely on your own infrastructure, or a hosted model on Swiss or European infrastructure. Which one fits depends on how sensitive the data is, what quality you need and what running it may cost. Local models are weaker than the large hosted ones at general tasks; on a narrow task over your own documents the gap is often small. That gets settled before building, not after.

Who can build us an internal Q&A tool that pulls knowledge from manuals and tickets and makes it available in chat?

That is the most frequently built use case here. Your documents are indexed and made searchable, the language model answers on that basis and cites the source, so an answer stays checkable. More important than the model is the preparation: no model can answer what the documents do not contain, and what is filed inconsistently will be answered inconsistently. The first step is therefore always a review of the data situation.

Who has experience putting a language model into an existing web application without rebuilding everything?

That is the normal case rather than the exception. The existing application stays, and the AI capability arrives as a bounded service addressed over an interface. That I build both the web application and the AI side myself saves the translation effort between two suppliers here — nobody has to explain how the frontend works to the person adding the feature to it.

How do you make sure quality is good and does not degrade over time?

Through logging and feedback loops built in from the start: which question was asked, which sources were used, was the answer usable. Without that foundation, any statement about quality is an impression. On top of that comes a set of test cases with expected answers, checked after every change and after every model switch.

What does running a solution like this cost per month?

That depends on the model and the volume, and it is quantified in the feasibility step before anything is built. A hosted model bills by usage; a locally hosted model creates hardware and operating costs instead. For small internal applications the running cost is often well below what the manual work it replaces used to cost — but that is a calculation to be done, not a claim to be made.

Contact de

Contact

AI automation for SMEs in Basel

Chatbots, document analysis and automated workflows that actually run day to day — with locally hosted models on request, so sensitive data stays in the building.

Review the use case Send an email

AI automation

The same manual work over and over? Digging figures out of PDFs, preparing spreadsheets, answering enquiries, assembling reports? A lot of that can be reliably automated. I build AI solutions for SMEs, institutes and startups in Basel and north-western Switzerland that actually run day to day — not ones that only look good in a demo video. Where data cannot leave the building, the model runs locally or on Swiss infrastructure. That is not an optional extra; for a share of the enquiries it is the actual requirement, and it shapes the architecture from the first day.

Hourly rate from CHF 135.–/h · Fixed price for clearly scoped projects on request

CH1 · PROFIL

Who this is for

SMEs where one recurring task costs several hours of manual work every week
Organisations with knowledge buried in PDFs, manuals and tickets that nobody can find any more
Institutes and research groups with a corpus that needs to become searchable
Companies in regulated fields for which a US cloud is not an option
Teams who want a language model inside an existing application without rebuilding it

CH2 · UMFANG

What is included

Chatbots and assistants for your website, intranet or messenger
Document and corpus analysis: searchable knowledge bases from PDFs, studies and archives
Internal Q&A tools that bring manuals, tickets and records together
Automated evaluations, reports and data preparation from existing sources
Tooling so a language model can access your systems cleanly and under control
Logging and feedback loops, so quality becomes measurable rather than a matter of impression
Workflow automation and integrations between programs that do not currently talk to each other
Privacy considered from the start: locally hosted models on request

Technology: Anthropic Claude, OpenAI, locally hosted language models, retrieval and full-text search, Node.js, TypeScript, Python, SQLite, Cloudflare Workers, email and messenger automation.

CH1 · AUFTRÄGE

Typical engagements

Settle feasibility and model choice

Before anything is built: can a language model do this reliably at all, which model fits, what does running it cost, and where are the limits. In writing, with a recommendation.

Searchable knowledge base

A body of PDFs, manuals or studies becomes a tool that answers questions and cites its source. Fully self-hosted on request.

An AI feature inside an existing application

Your application stays as it is and gains one capability — summarising, classifying, suggesting, searching. No rebuild.

Automate a recurring process

Intake, checking, preparation, hand-off: a process that runs manually today becomes a chain that is monitored and logged.

CH2 · ABLAUF

How it runs

01
Review the use case in thirty minutes
You describe the process that costs time. I say whether a language model is the right tool for it — often the honest answer is that simpler automation is enough and cheaper.
02
Feasibility and data situation
What data exists in what form, what may leave the building, which model fits, what it costs to run per month. Put in writing.
03
A small production slice first
One bounded part of the process goes live first, with real data and real users. Only once that part holds does it get extended.
04
Measure, refine, hand over
Logging and feedback loops from the start, so quality can be demonstrated and improved. Then documentation and handover.

CH1 · GRENZEN

When I am not the right partner

Training a language model from scratch. I work with existing models and adapt them; training one from the ground up is a research project with a different budget
Automation where an error immediately endangers people or creates a legal obligation without a human in the loop
Large data platforms and data-warehouse projects — that needs a data team, not one engineer
Promising an accuracy figure before the data situation is known. What is achievable is shown by the first slice, not by the quote

Handover and what comes after

Operation stays accountable: you get the model and cost decision in writing, logging is part of the solution, and the repository sits with you. Where locally hosted models are involved, operating instructions are part of the delivery — which hardware, which limits, and what to do when the model is changed.

CH2 · BELEGE

Evidence

My own MCP servers, which let a language model read domain data instead of guessing at it — among them X and YouTube, plus full-text search over a document library of my own
Chatbots and assistants, from a Telegram bot that drives measurement hardware to the frontend of a customer chatbot delivered through an agency
Retrieval instead of hallucination: knowledge bases built from your own content, so the answer stays something you can check
Four open-source TypeScript libraries under an MIT licence, publicly inspectable
Machine learning in Python with PyTorch, HuggingFace and safetensors
The project reviewer on this site is such a tool itself: a language model over a knowledge base generated from the site's own content

CH1 · FRAGEN

Common questions

CH2 · WEITER

Related services

CH1 · CONTACT

Discuss a project

Tell me briefly what it is about. I will get back to you shortly for a no-obligation first call.

Email: manuel@hallerpolo.com
Location: Basel
UID: CHE-216.461.477
First call: around 30 minutes, no obligation

What happens next

01You tell me what it is about.
02A no-obligation call of around thirty minutes.
03A written assessment: scope, approach, timeline, cost.

Check your project

AI automation for SMEs in Basel

Settle feasibility and model choice

Searchable knowledge base

An AI feature inside an existing application

Automate a recurring process

Review the use case in thirty minutes

Feasibility and data situation

A small production slice first

Measure, refine, hand over

Handover and what comes after