Commercial via API
Frontier models from major providers — where your data policy allows sending the data out.
AI Development · Enterprise AI
Enterprise AI systems and private LLMs
RAG over your knowledge, AI agents for your processes, assistants for your staff. They work with your data and your access rights, on commercial, open-weights or custom models — inside your perimeter when the data must not leave it.
Discuss your use caseThe model is only part of the system
An enterprise AI system is a language model connected to your data and processes: search over the knowledge base, tools that act in business systems, access rights, an audit log and quality checks. The model can be swapped in a day; how the system handles data and errors decides whether it is still useful a year later.
What we build
-
RAG over enterprise knowledge
Answers over your policies, procedures, contracts and technical documentation, with a citation to the source. We respect document permissions, not only document content.
-
AI agents for business processes
Agents that act, not only answer: look things up in business systems, prepare requests and documents, run checks. Tools connect through the open MCP standard; actions with consequences are confirmed by a person.
-
Assistants for staff and support
First-line support, an expert's co-pilot, an assistant in a messenger. Requests are routed between models by task type — a cheap model for the simple, a strong one for the hard.
-
Document extraction and classification
Turning incoming emails, requests, forms and scans into structured data for your systems, with quality checked on samples.
-
Private and on-premise LLMs
Models deployed on your hardware or in an isolated segment, so data stays inside your perimeter. Models and hardware are chosen by measurements on your tasks.
-
A governed AI layer for IGA
Our own example of an enterprise AI system: a governed AI agent for identity administrators.
Models
Commercial, open-weights and custom models
We choose the model by task and hosting constraints, based on measurements on your data. One system often runs several models, each on its own type of request.
Open-weights, private
Qwen, DeepSeek, gpt-oss and similar models running inside your perimeter. Data never reaches the model vendor.
Custom
Classifiers, retrieval models and compact models trained for one task — where a large model is overkill or too expensive.
How we measure quality
A test set of questions with reference answers, built with your experts. Retrieval — were the right passages found — and generation — does the answer rely on them — are measured separately. The set is re-run on every change of model, prompt or knowledge base, so a fix in one place does not break another.
How we contain risk
The system runs with the permissions of the user who asked: it sees only what that user can see. Secrets are masked before reaching the model, every request and action is logged, and operations with consequences are confirmed by a person.
Choosing the approach
RAG or fine-tuning
| Need | RAG | Fine-tuning |
|---|---|---|
| Knowledge changes often | Fits: update the base, answers update | Requires retraining |
| A source must be cited | Fits: answers rely on passages | No source to show |
| Narrow style, format, terminology | Partly, via instructions | Fits |
| A small model instead of a large one | Does not help | Fits: a compact model for one task |
Track record
Examples
No customer names — only the problem and how the system works.
A governed AI layer for identity administrators
Our governed AI layer for IGA: 11 read-only tools over the MCP standard, RAG over access policies, runs with the user's permissions, masks secrets. Vendor-neutral over midPoint, SailPoint and other IGA platforms, on the LLM you already operate.
An assistant with model routing
Our own customer-facing assistant in a messenger: different request types are served by different models — including Qwen, DeepSeek and gpt-oss — balancing answer quality against cost.
How we start
-
Discovery
We map the process, data and hosting constraints, and build a test set of questions and acceptance criteria with your experts.
-
Pilot on your data
A working system on your real knowledge base and tools; models compared on the test set; permissions and logging verified.
-
Production
Deployment in your environment, quality monitoring, scheduled test-set runs and knowledge-base updates.
Video analytics and object recognition are covered on Computer vision development.
FAQ
Frequently asked questions
Can we use LLMs without sending data outside the company?
Yes. We deploy the model inside your perimeter — on your own hardware or in an isolated segment — and connect it to data and systems inside it. Where data policy allows, a managed cloud model can be used for the less sensitive workloads.
Which models do you use?
Three groups: commercial models via API where your data policy allows it, open-weights models such as Qwen, DeepSeek and gpt-oss running inside your perimeter, and our own models trained for a specific task. The choice is made by measurements on your data: quality, speed, cost and hosting constraints.
What is RAG and when is it better than fine-tuning?
RAG means the model answers from passages retrieved from your knowledge base and cites them. Knowledge is updated without retraining and answers can be verified. Fine-tuning is for style, format or narrow terminology, not for fresh facts. The two are often combined.
How do you make sure the assistant does not make things up?
We build a test set of questions with reference answers together with your experts and measure retrieval and generation separately: were the right passages found, and does the answer rely on them. The set is re-run on every change of model, prompt or knowledge base.
What is the difference between an AI agent and a chatbot?
A chatbot answers questions. An agent executes process steps: calls tools, reads and writes data in business systems, checks the result. That is why an agent needs permissions, an action log and human confirmation for operations with consequences.
How does the EU AI Act affect an enterprise AI project?
It depends on the use case. Obligations for general-purpose AI model providers have applied since 2 August 2025. After the 2026 AI Omnibus, obligations for high-risk systems apply from 2 December 2027 for Annex III use cases and from 2 August 2028 for Annex I products. We design data flows, logging and human oversight so that your compliance team can assess the system; the legal qualification stays with them.