Twenty years of process documents, turned into an in-house expert that never leaves the intranet
An advanced equipment manufacturer
What it looked like before
The company had accumulated more than twenty years of process documents, equipment manuals and fault-handling records, and a new hire needed six months or more shadowing senior engineers. A public-cloud AI approach had been tried before, but process parameters are core assets — data leaving the intranet was unacceptable.
What we delivered
A full-stack private deployment of Uniclaw: a self-hosted Milvus vector store, a knowledge base loaded with company documents, and a privately deployed model service speaking the OpenAI-compatible protocol. Three-tier memory lets the agent remember how different roles ask and what they care about, and the boundary layer of the five-layer persona forces anything outside the knowledge base to a human.
- The vector store and model service both run inside the customer’s intranet; process parameters never cross the boundary
- Three-tier memory settles by role, so senior engineers’ experience remains in retrievable form
- The persona boundary layer states explicitly “what must be handed to a human”, keeping the AI from guessing on critical process questions
- Answers carry source citations, so engineers can follow them back to the original documents
We are not trying to replace the senior engineers — we want their experience to be askable at three in the morning.
Hand the complexity of identity, agents and private domain to one governable kernel
Whether you are replacing an existing IAM, building an agent platform, or trying to make private-domain operations actually work — start with a 30-minute architecture call. We will first judge whether this is the kind of problem we are good at, and say so plainly if it is not.