Skip to main content

AI development

AI development means building solutions on top of language models that survive daily use. The demo is the easy part. I have been a developer since 2010 and write the code in every project myself. We start small, measure the result and expand from there.

  • I usually reply the same day
  • Written estimate before work starts
Mikkel Tschentscher

star reviews

Read every review
  • Helpful and lightning fast to reply

    Mikkel is seriously good, helpful, lightning fast to reply, and on top of that he is a nice person. I can warmly recommend Mikkel to others..

    Trine Handskemager(translated from Danish)
  • A fast, trouble-free process

    Mikkel delivered competent, professional work setting up my website. Everything ran quickly and without a hitch.

    Sisse Find(translated from Danish)
  • My go-to guy

    That is why I can still give Mikkel my very best recommendations.

    Janni E Clausen (Janniclausen)(translated from Danish)

AI development

From prototype to production

A prototype only has to convince in one meeting. A solution in production has to handle the inputs nobody predicted, respond within reasonable time, cost the same next month and fail in a way you can detect.

That distance is larger than it looks, and it is usually where I get brought in. Either to build from scratch with production in mind, or to make a promising prototype robust enough for real users.

Typical projects

  • Customer service: suggested replies, case summaries and search across previous enquiries.
  • Document handling: extracting data from contracts, invoices and technical documents.
  • Internal tools: search and answers on top of your own documentation, with sources you can click through to.
  • Integration with existing systems: function calling against your CRM, accounting system or own API, so the model works with real data instead of guessing.

Evaluation and measurable quality

Without an evaluation setup you cannot know whether a change made the solution better or worse. You only have a feeling, and feelings are unreliable when the answer differs every time.

So I build a test set of real examples with expected answers early, and measure against it every time the prompt or model changes. It is the single thing most often missing in the projects I take over from others.

What is AI development?

AI development is software development where a language model forms part of the solution. That might be search across your own documents, automatic handling of incoming enquiries, or a tool that pulls structured data out of PDFs. The model provides the language understanding, while ordinary code takes care of data access, security and integrations with the systems you already run.

The need typically arises in companies with a lot of text and a lot of repetition. A customer service team answering the same questions over and over, a finance team keying in invoices by hand, or a department spending hours hunting for answers in internal documents. If that kind of work exists in your business, there is probably something to gain.

There is a big difference between being able to talk about AI and being able to build it. An advisor can point at possibilities, but only once someone writes the code does it become clear what your data can actually support and where the limits are. I am a developer with 16 years of experience, and my recommendations come from things I have built and put into production myself.

My way of working is the same in almost every project: we find the smallest version of the solution that can create value, build it in a few weeks and measure the result against a test set. You own the code and the prompts from day one, and there is no lock-in to me or to any particular vendor.

What the pace of AI development means for businesses

Artificial intelligence is moving fast, and it is easy to be unsure when to get started. New language models arrive every few months, they get better at following instructions, reading long documents and using tools, and the same task typically gets cheaper to run over time. That can make it seem sensible to wait for the next generation before building anything.

In practice it is rarely the model that decides whether an AI solution works. It is your data, the integrations with the systems you use, and a test set that shows whether the answers are good enough. That work does not go out of date when a new model arrives. On the contrary, it is exactly that work which makes it possible to switch to a new model without rebuilding and see straight away whether it is better.

So my advice is to get started, but to build loosely coupled. Generative AI keeps moving the line for what is possible, and once you have an AI solution in production, you can adopt each new model without rebuilding. If you wait for the perfect model, you never get started, because there is always a new one on the way. It does require that quality is measured, so an upgrade becomes a decision based on numbers rather than on a press release.

Open-source or closed models? Cloud or local servers?

Closed models, like the ones behind ChatGPT and Claude, are used through the vendor's API. You pay for usage, have nothing to run yourselves and usually get the strongest models on the market. On the other hand, data is sent to the vendor, so the data processing agreement has to be in place: where the data is processed, how long it is stored, and whether it may be used for training.

Open-source models, or more precisely open-weight models, can be downloaded and run yourselves, on your own servers or with a hosting provider of your choice. Data does not leave the boundaries you set, and you are not dependent on a vendor's price changes or on a model being retired. The downside is that you have to handle hardware, operations, updates and security yourselves, and that open models often lag a little behind the best closed ones on the hardest tasks.

My starting point is closed models via API, most often Claude or OpenAI, with EU hosting or agreements where data is not used for training. That covers most tasks and requires no operations on your side. A local model makes sense when data must not leave the building at all, or when volume is so large that running it yourselves pays off. Because I build with a thin layer between the code and the model, the choice can be revisited later without starting over.

AI risks: hallucination, bias and confidential data

A language model can answer wrongly with great confidence. This is called hallucination, and it happens because the model predicts a likely answer instead of looking something up in a source. When I build a solution that answers questions, it therefore retrieves the relevant passages from your own documents, lets the model answer from them and shows the source next to the answer. If it finds nothing, it should say so instead of guessing, and the test set shows how often it still goes wrong.

Bias means the model repeats skews from the data it was trained on, for example in the way it describes or assesses people. For tasks like sorting invoices it rarely matters. If the solution is to assess applicants, customers or employees, on the other hand, it has to be tested on examples with broad coverage, and a person has to make the decision. Uses like these may also fall under the stricter requirements of the EU AI Act.

Confidential data is the risk that is easiest to overlook. An internal search solution must not show an employee documents they do not already have access to, so the access control from your existing systems has to carry over into the solution. I keep personal data out wherever possible, and where that is not possible, I document where the data ends up. It is dull work, but it is what makes the solution trustworthy.

AI development for companies across Denmark

I am based in Glostrup just outside Copenhagen, but I build AI solutions for companies in Aarhus, Odense, Aalborg and several smaller towns, plus a few clients abroad. Most of them found me online, and several have never been in the same room as me.

Development work suits distance well. Meetings run over video, the code lives in your repository from the start, and status arrives in writing as we go, so you can always see where the project stands. If you are in the Copenhagen area, I am happy to come by for kickoff and the important decisions.

Typical projects are a search solution on top of internal documentation, an integration between a language model and a line-of-business system, or an internal tool that removes manual data entry for a whole team.

As a freelancer I am the same person all the way through: whoever estimates the job is also the one writing the code, and there is no middle layer when you have a question. An agency can field more people, but in AI projects of this size, continuity is usually worth more.

When AI is the wrong choice

A share of the ideas put in front of me would be better solved with ordinary software. If something needs classifying by three fixed rules, three if-statements are cheaper and faster, and anyone can read them. A language model adds uncertainty in a place where none needs to exist.

Volume is the other thing I look at. If a task happens fifteen times a month and takes ten minutes each time, there is an hour and a half to save, and that rarely carries a development project. The arithmetic has to work before I recommend building.

I say so out loud when it does not, even when it costs me the job. In return, you can trust the recommendation when I say something is worth building.

What you get

Concrete deliverables, not a report of recommendations someone else has to find time for.

  • A working system

    A solution running in your own infrastructure that handles errors and can be monitored in production.

  • An evaluation set

    A test set of real examples with expected answers, so you can measure for yourselves whether a change improves quality.

  • A running-cost estimate

    A calculation of what the solution will cost to run at your volume, done before we build.

  • Documentation

    A description of how the solution fits together and what to do when things go wrong, written for whoever takes over.

How it works

Four steps. You know what happens when, and you can stop after any of them.

  1. 01 / 04

    Pick the process

    Together we find the piece of work that repeats often enough for automation to pay off. The more contained the task, the faster the first version reaches production.

  2. 02 / 04

    Set a baseline

    We record how long the task takes today, how often it occurs, and how many errors slip through. Without those numbers, nobody can tell afterwards whether the solution was worth the effort.

  3. 03 / 04

    Build against a test set

    I collect real examples from your day-to-day and test every prompt and model change against them. That keeps quality stable when the solution is tuned later on.

  4. 04 / 04

    Production with logs and a kill switch

    The solution is wired into your existing systems with full logging and a switch that can turn it off. I document the setup and train someone on your side to maintain it.

Brands I've worked with

  • MT Højgaard Danmark
  • Egmont
  • NNIT
  • Visma
  • Lomax
  • Pascal
  • able.
  • OOONO
  • Novo Nordisk Fonden
  • Energii
  • Maersk Tankers
  • Nordkysten Entreprenørfirmaet

Frequently asked questions

How does AI development differ from ordinary software development?

The foundation is the same, but a language model does not answer identically every time, and that changes how you test and handle errors. Evaluation and monitoring therefore take up far more room than in a classic project. The craft of integrations and data models is familiar territory.

What can AI be used for in a business?

Mostly for work with a lot of text and a lot of repetition. Typical examples are search across your own documents, sorting incoming enquiries with suggested replies, extracting data from invoices and contracts, and drafting texts that someone on your team edits and approves. What they have in common is that the task happens often and follows a pattern.

What is the difference between AI and ChatGPT?

AI is the broad term for systems that can do tasks that normally need human judgement. ChatGPT is one product built on a language model from OpenAI. When I develop AI solutions, I use the same kinds of models, but built into your own systems and data, so they solve a specific task instead of being a chat someone has to remember to open.

What is generative AI?

Generative AI is artificial intelligence that can create new content such as text, code or images from an instruction. The language models behind ChatGPT and Claude are the best-known examples. In AI development they are used just as often to read and understand text, for example to sort enquiries or pull data out of documents, as to write something new.

What is hallucination in AI?

Hallucination is when a language model gives an answer that sounds convincing but is wrong or made up. It happens because the model predicts the most likely answer instead of looking something up in a source. The risk cannot be removed entirely, but it shrinks considerably when the model has to answer from your own documents, show the source and say so when it does not know.

What does AI development cost?

It depends on the scope. The three things that drive the price most are the number of systems to integrate, the quality of your data and the precision requirements. Describe the task in an email and I will ask a few clarifying questions and come back with a concrete quote, with a reply the same day.

Which models do you work with?

Mainly Claude and OpenAI, but I always build with a thin layer in between so the model can be swapped later. Price and quality in this field change fast enough that locking into one vendor is a bad idea.

Should we use open-source models or the models behind ChatGPT and Claude?

For most companies, a closed model via API is the best place to start: it is strong, needs no operations, and with EU hosting or an agreement that data is not used for training, it can be used for most tasks. An open-source model on your own servers is worth considering when data must not leave the building at all, or volume is very large. The choice is not final, because I build so the model can be swapped.

How do you handle GDPR and data security?

We start by clarifying what data the solution actually needs, because personal data can often be kept out entirely. Where it has to be included, I use EU-hosted models or agreements where data is not used for training, and I document the processing so your DPO or advisor can assess it.

Do you only work with companies in Copenhagen?

No. I live in Glostrup outside Copenhagen, but a large share of my clients are elsewhere in the country, including Aarhus, Odense and Aalborg, and a few abroad. Development works fine over video and in writing.

How long before something is in production?

A bounded solution on one system is typically four to eight weeks including testing. If data has to come from several systems, the integrations set the pace, not the AI part.

Who owns the code and the prompts?

You do, from the first commit. The code lives in your repository, the prompts are documented, and you can take everything to another developer at any time. I do not build anything that leaves you dependent on me.

Who maintains the solution afterwards?

We agree on that before we start. If your own team can take over, I build it so they can and write the documentation for them. If they cannot, we set up an ongoing arrangement instead of leaving the question hanging.

Your next step

Get a quote for AI development

Tell me about your project. I’ll get back to you with questions, options and a concrete next step.