An internal AI platform where staff build their own agents
A proof of concept for a 10,000-person enterprise, run with its 300-person IT department: every model in one chat, a personal document library, the company's policies answerable, and two agent builders, one visual for engineers and one drag-and-drop for everyone else.
Context
An enterprise of more than 10,000 people wanted to know what an internal AI platform should be before building one for everyone: a place where staff could bring their own documents, ask about the company’s own procedures, and put together agents on the company’s own data without filing a ticket. The proof of concept ran inside the IT department, about 300 people, to find out what they would actually build.
Constraints
- Confidential documents had to stay inside: uploaded to a personal library and usable within the owner’s area, never pooled.
- Every available model behind one interface, so choosing a model was a setting rather than a purchase.
- Agents could only reach a fixed catalogue of tools. New tools were added after review.
- Two kinds of builder. Engineers wanted branches, tools and documents wired visually; everyone else needed something they could assemble in minutes.
- Twenty contractors building in parallel, so the platform had to be services with clear seams from the first week.
What we built
A web platform of microservices, React and Node at the front, Python services behind, with one chat over every model available on Vertex AI. Each person had a library for their own documents, confidential ones included, searchable within their area. A company knowledge base held the procedures, guidelines and policies, so “what is the working-from-home policy” got an answer with the source. Documents could also be turned into a podcast for people who would rather listen.
The two agent builders mattered most. The advanced one let an engineer configure an agent visually: the tools it may call, the documents it may read, the branches it takes based on its own decisions. The front end wrote that configuration as YAML and the back end turned it into a LangGraph agent on the fly. The simple one was drag and drop for everyone: put tools and documents in a folder, write a prompt saying how to use them, done. Tools were MCP servers from a fixed list, giving access to BigQuery, live parcel data, documents and more, with the data itself in Spanner and BigQuery.
What people built with it answered the question the proof of concept was asked: agents that pulled data from BigQuery and live parcel data to compare the two, and agents that decided which tasks to generate from what they found.
Outcome
A working platform with 300 users, and the questions a production rollout needs answered: which tools to expose, how documents are scoped, and what non-engineers build when given a builder they can use. Jakub designed the platform, built the proof of concept and led the project; a team of twenty contractors delivered the services.
Stack
React, Node, Python, LangGraph, MCP servers, Vertex AI, Spanner, BigQuery.