A custom AI model for your organization: from data to deployment
General-purpose AI models do many things, but they do not know your organization's knowledge, internal documents or specialist terminology. In many organizations, sending data to external services is also impossible or not allowed. The solution is a model that works with the organization's own data and runs on the organization's own server.
You don't always need to train a model from scratch
- Retrieval-augmented generation (RAG): when answering, the model finds relevant passages in the organization's documents and answers based on them. For answering questions from regulations, technical documentation or records, this is usually the fastest and cheapest approach, and answers stay current as documents are updated.
- Fine-tuning: a ready-made model is trained further on the organization's examples so it handles a tone, an output format or a specific task (such as classifying letters) better.
- Training from scratch: unnecessary for most organizations, and it requires very large amounts of data and computing resources.
Data: the most important part of the project
A model is only as good as its data. Before starting, define exactly what the model must do and how its success will be measured. Data is then collected, cleaned and labeled where needed. Personal and confidential data must be reviewed before use and anonymized where necessary.
Running on your own server
Open language models are released in various sizes. Small and medium models can run on an in-house server with one or more GPUs, without any data leaving the organization. GPU memory (VRAM) usually determines which model can run and how fast; techniques such as quantization reduce the memory needed.
Example: intelligent analysis of camera footage
In an existing video surveillance system, a computer vision model can recognize license plates, count traffic or raise an alert when someone enters a restricted area, using the footage from the same cameras. The cameras don't need replacing; processing runs on a server next to the recording system.
Project stages
- Define the task and the success criteria
- Review and prepare the data
- Build a prototype and evaluate it on real data
- Deploy on the organization's server and connect it to existing software and equipment
- Monitor performance and update