AI integration
AI integration means connecting a language model to the documents, databases and tools your company already has, so it answers with your information instead of general knowledge. We build internal assistants, customer-facing AI features and agents that take actions in your systems, with tests that show how often the answers are right.
An example
An internal assistant answers only from the company’s own documents and cites them. When the documents don’t say, it says so.
What we build
- Assistants over company documents
- Answers staff questions from policies, contracts, manuals and wikis, with a link to the source for every answer.
- AI features in your product
- Summaries, smart search, drafting and classification added to your SaaS or internal tool, with usage limits and cost tracking.
- Agents that use your tools
- Assistants that can look up a customer, create a ticket or update a record, with permissions you control.
- Private model hosting
- Open-weight models on your servers or a private cloud when data cannot leave your environment.
How a project runs
- Use case and data review. We look at the questions people ask, the documents that hold the answers and who is allowed to see what.
- Evaluation set first. Before building, we collect real questions with correct answers. Every version is scored against them.
- Build and connect. Ingestion, retrieval, prompts, permissions and the interface: Slack, Teams, a web app or your own product.
- Measure and improve. We track answer quality, cost per question and the questions it could not answer, and improve from there.
What to have ready
- Where your documents and data live (Drive, SharePoint, Notion, databases)
- A list of 20 to 50 questions people really ask
- Who should see which documents
- Constraints on where data can be processed
Tools we use: Claude, OpenAI, Llama and Qwen (self-hosted), pgvector, Qdrant, LangGraph, Python, TypeScript
Questions we get asked
Do we need to train our own model?
Rarely. Most business cases work best with an existing model plus retrieval from your documents (RAG). Fine-tuning helps with tone or format, not with knowing your latest data.
How do you stop the AI from making things up?
Answers are built from retrieved passages and cite them. When the documents do not contain an answer, the assistant says so. We measure this on an evaluation set before launch.
Which model do you use?
It depends on the task, language, cost and data rules. We commonly use Claude and OpenAI models through their APIs, and open-weight models such as Llama or Qwen when hosting must be private.
Can it respect document permissions?
Yes. Retrieval is filtered by the user’s access rights, so people only get answers from documents they are allowed to read.
Tell us what you need
A few sentences are enough. We reply by email with questions or a time for a call.
Other services: AI voice agents, AI automation, Web and SaaS development, Mobile app development, Blockchain development