AI and local models
We choose the model according to the task and run it where your data is allowed to go: on your server or ours, or through an API without training on your data. Answers have a source, and a person decides on irreversible steps.
What we can do
We build work with AI from parts we know in detail. Each one is described by how it works and, with an example, what it does.
-
Local models
The model runs on your server or ours and the data never leaves the infrastructure. A server without a graphics card is enough for search by meaning, while larger models run on servers with a GPU.
Task search by meaning on this website result model on our server, without a GPU
-
Models through an API
For less sensitive tasks we use the provider's API in a mode without training on your data. You decide where the data may go.
Task summary of public texts result API without training on your data
-
Choosing and swapping the model
On a sample of your data we compare several models in answers, speed and cost. The application talks to the model through a unified interface, so the model can be replaced later.
Situation a different model provider result configuration change, application unchanged
-
Search by meaning
We convert text into a vector, and a query finds a document by meaning even if it does not contain the same words. If the model does not respond, search carries on with full-text.
Query connection to artificial intelligence result a page about AI in a company, with no matching words
-
Answers with sources
The assistant answers from your documents and shows what each answer is based on. It sees only the data the user is authorised to access.
Query a new colleague's query in their own words result an answer with a link to the document
-
MCP servers
We build an MCP server through which an AI assistant reaches your systems and data. It sees only what the user is authorised to access, and a person confirms irreversible actions.
Query assistant query about order status result an answer straight from your system
-
Text, image and sound
From free text, emails, photos and scans it reads data like a person would, and it transcribes and summarises calls and meetings. It also handles languages other than the one you ask in.
Source free text of an auction notice result data in the property record
-
Answer checking
We store what the model received and what it returned, and protect it against prompt injection. After a model swap or a prompt change, we run the answers through test examples again.
Situation prompt change result answers again on test examples
-
Costs under control
The pilot shows the cost of one query in operation. Consumption has a cap and repeated queries come from the cache, so they are not computed again.
Situation repeated query on this website result from the cache, without a new calculation
How we build it
We start with the task, not the model. First a small pilot on a sample of your data; the scope is decided by the result. We will also tell you where you do not need AI.
-
Task and data What the AI should do, on which data, and how we tell that it works well. Without the third point there is no sense in starting.
-
Pilot and model comparison On a sample of your data we try several variants and show the difference in answers, speed and cost.
-
Deployment and operation Integration into the system, usage limits, query logging and continuous tuning based on actual operation.
Technology
- Local models
- Ollama, open models such as Gemma or Qwen
- Models through an API
- ChatGPT, Claude, Gemini, in a mode without training on your data
- Integration into the application
- Symfony AI, Node.js, Python
- MCP servers
- connecting AI assistants to your systems and data
- Vectors and search
- PostgreSQL with pgvector, HNSW index
- Servers
- servers with or without a GPU, private instances at the provider
This is a selection. We choose the model according to the task, data sensitivity and budget, and it can be replaced without rewriting the application.
Where models work for us
AI reads data from free text and helps in the administration. On this website it searches by meaning on our server.
We can do a lot
A model is only as good as the data and search beneath it. We build every project from these capabilities.
- System integration and data collection Systems from different vendors working as one. Adol Monitor reads dozens of sources every day.
- Advanced data search Typos and meaning, hundreds of thousands of records, responses in tens of milliseconds.
- Servers and operations From a shared server to dedicated servers and the cloud. Designed, run and monitored 24/7.
- Speed and performance A website that loads in the blink of an eye, and a portal for tens of thousands of people a day.
- Maps and geodata Addresses in the right place, nearby search and route planning.
- Findability and analytics Visible in search engines and to AI assistants. Measuring your goals, even without cookies.
- Cookies, GDPR and accessibility A consent bar by the rules and, where possible, a website without one. Accessibility from the start.