
Semantic search and intent-based systems
What is semantic search? How vector and hybrid search work in Meilisearch, what they do for online shops and chatbots, and how we build it in Laravel.
We use the data you already have for forecasts, automation and better decisions, and run the models behind stable APIs connected to your systems.
We build and run systems for
We apply algorithms to concrete problems in your business, and build the system around the model so it actually gets used.
Historical data gives forecasts for stock, sales and maintenance, so you can plan ahead.
Natural language processing (NLP) sorts incoming requests and documents, which saves your staff a lot of manual work.
Algorithms that look at how people use your service and show each user content and products that fit.
Computer vision that checks quality on the production line, or that reads and interprets scanned paper documents.
Models that spot unusual patterns in large datasets, such as attempted fraud, system faults and security breaches.
Open language models on your own servers, fine-tuned on your terminology, with no data leaving your servers.
Machine learning does the most when the problem keeps coming back and the data already exists.
Plenty of historySales, orders or measurements going back several years.
Recurring judgementsCases and documents that are sorted by hand today.
Patterns nobody seesAnomalies and connections hidden in large amounts of data.
Sensitive dataInformation that cannot be sent to an external AI service.
Embeddings from machine-learning models let search understand what users mean. This is how we build hybrid search in Meilisearch.

What is semantic search? How vector and hybrid search work in Meilisearch, what they do for online shops and chatbots, and how we build it in Laravel.
We build machine-learning systems in Python and put them into production alongside the business systems you already have. Machine learning lets a program learn patterns from historical data instead of a developer writing every rule by hand, and Python is the most widely used language for it. To us, sound system architecture matters as much as the algorithms, because that's what turns your business data into forecasts, automation and better decisions.
We're software developers at heart. A machine-learning model is no use while it only lives in a Jupyter notebook. We put models into production, wrap them in stable APIs and connect them to your existing systems.
Artificial intelligence is a broad field that covers everything from rule-based systems to large language models. Machine learning is one part of it. Instead of someone writing rules like "customers who haven't bought anything in 90 days are at risk", we show the model thousands of examples of customers who left and customers who stayed. The model works out for itself which signals matter and how much weight each one carries.
Most machine learning in business comes down to three kinds of task:
Large language models like ChatGPT are machine learning too, but someone else has trained them on enormous amounts of text. On this page we usually mean models we train on your own data to answer your own questions.
Machine learning fits when a problem keeps coming back, the answer can be learned from history and the data already exists. These are the tasks we work on most often:
Machine learning isn't always the right tool, and we'll tell you when we think so. These are the most common reasons to choose something else:
If you want AI to answer questions from your own documents or carry out tasks in other systems, a language model with search is often a better starting point than a model of your own. We cover that under AI and automation.
It depends on the problem you want to solve. For prediction you need historical data with at least a few hundred data points. For classification you need examples of each category, ideally labelled by someone who knows the subject.
Many businesses have more data than they think, spread across the accounting system, the CRM, the support tool and the logs. The first thing we do is check whether the data is good enough. The answer may well be no, and it's better to find that out after two weeks than after two months.
The technology we pick depends on three things. The first is the kind of data you have: structured tables, or text, images and audio. The second is how precise the answer has to be, and the third is how important it is to explain why the model answered the way it did. No single setup fits every case.
We know the Python ecosystem well and choose libraries based on the technical requirements. A simple model you can explain is often worth more than a heavy model that's slightly more accurate.
For classic machine learning on tabular data, such as spreadsheets, SQL databases and CSV files, scikit-learn is our workhorse. It has efficient algorithms for regression, classification and clustering, and it's usually where we start.
When we need more precision we use gradient boosting with XGBoost or LightGBM, and for classic statistical modelling we use Statsmodels. These models are fast and easy to interpret. We can often show exactly which variables carry the most weight, which you need when someone has to check the result.
When the data is unstructured, such as text, images or audio, you need deep neural networks. For that we use TensorFlow or PyTorch, which Google and Meta also build on, and Hugging Face Transformers for language models.
We build and train models that can see and read. Training takes a lot of computing power, so we optimise the code for GPUs. Running the model, known as inference, still has to be fast enough to answer in real time.
Many machine-learning projects never get further than a notebook on someone's laptop. The model works in a demo, but nobody uses it day to day.
We wrap the model in its own service, usually with FastAPI, that other systems can send requests to. Heavy jobs, such as retraining or running predictions on large volumes of data, run in the background with Celery, and Redis caches finished answers. The service runs in Docker and Kubernetes and connects to the web applications, business systems or Laravel applications you already have.
A model has a shelf life. When the market changes, or customers start behaving differently, its accuracy gradually drops. This is called model drift. We have long experience with DevOps, and we run machine learning through the same CI/CD pipeline as the rest of the code. That's what MLOps means:
OpenAI and others offer powerful APIs, but sensitive data often can't leave the building. We set up and run open language models such as Llama 3, Mistral and Gemma locally in your own infrastructure.
With tools such as Ollama and dedicated inference hardware, we can deliver AI features like chat assistants and document analysis that run entirely offline. That makes it much easier to stay within GDPR, and you avoid paying per token. We can also fine-tune the models on your terminology and your own data.
A simple proof of concept can be ready in 2 to 4 weeks. A production-ready system with integrations, monitoring and MLOps usually takes 2 to 4 months, depending on how complex it is.
A simple predictive model starts at around NOK 80,000. Larger projects with deep learning, dedicated infrastructure and ongoing maintenance are priced by agreement. Our hourly rate starts at NOK 1,500, and when the scope is clear you get a fixed price. The price depends mostly on:
Our AI projects go through the same four steps:
Everyone on our team has more than ten years of experience, and another senior developer reads the code before it goes to production. To see machine learning in use, read how we build semantic search.
These are the tools we use in most machine-learning projects, layer by layer.
Have data you think could do more for you? Book a chat and we’ll look at it together.
We check whether you have enough history and whether the data is clean enough to learn from.
We try several algorithms and pick the one with the best balance of precision and performance.
The model gets an API and is connected to the business systems you already use.
We track accuracy and retrain the model when new data arrives.
ongoingTypical examples are sales and stock forecasts, sorting requests and documents, recommendations, quality control with computer vision, and spotting fraud and system faults. What they have in common is that the problem keeps coming back and the data already exists.
When the rules are few and known, when you have too little data, or when every single answer has to be justified and a mistake is costly. Then ordinary code or a person is often the better choice, and we’ll tell you so.
It depends on the problem. For prediction you need historical data with at least a few hundred data points. For classification you need examples of each category. We help you assess whether your data is sufficient.
Artificial intelligence is a broad field that spans everything from rule-based systems to large language models. Machine learning is one part of the field, where the system learns patterns from data instead of following fixed rules. We use both, depending on what the problem needs.
Yes. We set up open language models such as Llama 3, Mistral and Gemma in your own infrastructure. The data never leaves your servers, which makes privacy and GDPR far easier to handle.
A simple proof of concept can be ready in 2 to 4 weeks. A production-ready system with integration, monitoring and MLOps usually takes 2 to 4 months, depending on complexity.
It depends on how complex the task is and how much data is involved. A simple predictive model starts at around NOK 80,000. Larger projects with deep learning, dedicated infrastructure and ongoing maintenance are priced by agreement. Our hourly rate starts at NOK 1,500.
Yes. We wrap the model in an API, usually with FastAPI, and connect it to your business systems, websites or Laravel applications.
We set up monitoring that alerts when accuracy falls below a threshold, and set up automatic retraining when new data comes in.
Tell us about the data problem you have in mind. We can help from the first idea until the model is running in production.