I build AI solutions

On open-weight AI models.

How I can help you

I can guide you through the whole process of deploying a local AI solution built on open-weight AI models (such as Qwen, Gemma, DeepSeek, Nemotron…) and hardware running right on your premises (a GPU server, an AI workstation…).

You decide the scope of our collaboration: it can be just one or a few consultations, or a long-term collaboration on rolling out the whole system.

A long-term collaboration can look like this, for example:

  1. Initial consultation

    We discuss how you want to use AI and whether a local AI solution makes sense for you. The initial consultation is free of charge.

  2. Solution design

    Within 1–3 weeks, I prepare a strategy for the solution. I present suitable hardware, models and options for connecting your tools.

  3. Deploying AI models

    I set up an inference server with AI models, connect your tools to it and prepare the documentation.

  4. Team training

    I teach your team to work with the new solution. I explain its benefits, limits, risks and possibilities.

  5. Maintenance

    I remain available to update the models, scale the hardware or implement new tools.

Why local AI?

Running AI models locally is not for everyone. The upfront investment in your own hardware can be high, and running the models takes more effort than with the usual commercial services. It is the right choice when any of the following applies to you:

  • Your data must stay private, with a guarantee.

  • Buying hardware works out cheaper for you in the long run than paying for a large volume of tokens.

  • Your AI solution must work reliably without censorship and without external dependencies.

  • Your AI solution must work offline too, on end devices.

What can you use open-weight AI models for?

For almost anything a commercial AI can do, there is an open alternative. Since you will be running these models on your own hardware, though, you need to consider the hardware demands of a model that can handle the task.

Here are a few examples of suitable tasks, along with examples of open-weight models that suit them:

How can local AI help your company or institution?

Concrete examples of how open-weight AI models can make your work more efficient:

Accounting firm

  • Sensitive data

Archive and library

  • Sensitive data
  • High volumes

Bank

  • Sensitive data
  • High volumes
  • Reliability
  • Data extraction Copies details from statements or identity documents into a table for further processing.
  • Automated reports Writes an automated report from client data (the generation can be rerun for an auditor at any time).
  • Assistant / chatbot Answers customers' questions about fees, loans and the like, following current policies and the price list.
  • Coding agent Helps update the banking system's legacy code, without the code ever leaving the bank's servers.
  • Speech transcription Transcribes call centre calls so that operators' conduct can be reviewed.

E-shop

  • High volumes

Energy

  • Sensitive data
  • Reliability
  • No internet

Hospital

  • Sensitive data
  • Reliability
  • Data extraction Pulls diagnosis and procedure codes out of medical reports for billing health insurers.
  • Translation Translates reports and instructions for patients who don't speak the local language.
  • Automated reports Drafts a discharge summary from the hospitalization records.
  • Assistant / chatbot Answers nurses' and doctors' questions about specific regulations and procedures.
  • Speech transcription Turns a recording of a conversation with a patient into an entry in the medical records.

Law firm

  • Sensitive data
  • Sorting and routing Matches incoming mail to the right case file.
  • Data extraction Copies terms from a contract into a table (e.g. change-of-control clauses, non-competes, contractual penalties).
  • Proofreading Checks a written document not only for grammar mistakes but also for inconsistencies and factual errors.
  • Document drafts Prepares the first version of a contract, a lawsuit or a statement from your templates and fills in details from the case file.
  • Assistant / chatbot Answers a legal question from case law or earlier filings.

Manufacturing company

  • Sensitive data
  • Reliability
  • No internet

Military and defense

  • Sensitive data
  • No internet

Police

  • Sensitive data

Public authority

  • Sensitive data
  • Sorting and routing Sorts incoming filings (applications, appeals, complaints) and passes them on to the right department.
  • Proofreading Finds references to wrong sections and inconsistently used terms in a draft regulation.
  • Summarizing long inputs Rewrites an official letter so that the recipient understands what to do without needing a lawyer.
  • Assistant / chatbot Answers civil servants' questions from internal guidelines and earlier case files.
  • Speech transcription Transcribes committee or council meetings.

Software company

  • Sensitive data
  • High volumes
  • Reliability
  • Sorting and routing Sorts customer requests and passes them on to the right team.
  • Automated reports Writes a report on bugs or security incidents from the logs.
  • Assistant / chatbot Answers developers and support staff from internal documentation, the wiki and resolved tickets.
  • Coding agent Generates code that never leaves the company. It can also help with cybersecurity tasks that commercial services often refuse.
  • Image generation Designs icons, illustrations and screen mock-ups for the product.

What we will work on

When deploying a local AI solution, we will work on these areas:

  1. Data

    We go through what data you have and where it lives: documents, databases, emails, scans. I propose how to convert it into a format the models can work with, and how to keep it up to date.

  2. Type of operation

    AI "with guarantees" can actually take several forms. Depending on how sensitive your data is, its volume and your budget, I recommend the one that suits you best.

    1. Fully local

      The models run on hardware you own, in your network. Data never leaves the building and everything works without the internet, too.

    2. Rented compute

      Open-weight models on a rented GPU server, for example in a local data center. You don't buy the hardware, but the server is yours alone.

    3. Open-weight models, paid per token

      Open-weight models through a provider's API, paying only for what you use. You can switch to another provider or to your own hardware at any time.

    4. Business plan of a commercial service

      Business editions of commercial services, with a data processing agreement and a guarantee that the provider won't train models on your data.

    More guarantees Easier to use

  3. Models, hardware and scaling

    We pick suitable models for your tasks and test them on your data. I work out what hardware you need for the output quality and request volume you require, and how that compares with paying for an API. I also show how to extend the solution once the request volume grows or a stronger model is needed.

  4. Software

    I design an interface on top of the AI model and connect it to your existing systems. I build tools for specific tasks, such as searching your documents.

  5. Maintenance and further development

    We plan regular maintenance of the system, model updates and further software development.

Why me?

  • I hold a PhD in language model research and have years of experience in the field.
  • I'm a freelancer: dealing with me is more straightforward than with a large company.
  • I use AI models every day in my own work.
  • I write a blog and a newsletter about open-weight AI models (in Czech) at Lokální.AI : Lokální.AI

I am a self-employed person operating under the Czech Trade Licensing Act. Company ID (IČO): 06845762. I am not a VAT payer.

Let's connect