Chat:R, an enterprise language-model platform
Published:
Role: Product and delivery lead. Client: Rud Pedersen Group. Timeline: 2025 to present.
Chat:R is an enterprise language-model platform I designed and led the delivery of for Rud Pedersen Group, a European public-affairs and strategic-communications consultancy. It serves approximately 1,000 staff across three Azure environments.
The problem
The firm’s consultants had begun pasting client briefings, policy positions and personally identifying material into public language-model tools, which created an unacceptable privacy exposure. Off-the-shelf alternatives were each unsuitable for a different reason: ChatGPT Enterprise resolved data residency but locked the firm to a single model, an awkward per-seat commercial structure at roughly a thousand staff and a feature set the firm could not extend on its own terms. What was needed was a ChatGPT-shaped interface deployed inside the firm’s Microsoft tenant, extensible to its working patterns and held as an owned asset rather than a rented subscription.
What was built
A bespoke React, Node.js and MongoDB application deployed on Azure, with three custom enterprise features:
- Projects, a workspace layer that groups conversations, agents and files by client or brief, modelled as a first-class MongoDB entity with REST endpoints and a virtualised React sidebar.
- Project-to-agent linking, an agent picker that surfaces each agent’s provider, model, tool list and file scope, so that specialised agents can be attached to a project without the consultant having to leave it.
- SharePoint MCP gateway, a Node.js Model Context Protocol server exposing a curated set of Microsoft Graph tools over a streamable HTTP transport, with credentials injected from Azure Key Vault.
Engineering decisions
- Virtual machines over managed Kubernetes. A cost and operational review after the first environment was stood up led to a considered pivot to Azure VMs. Monthly infrastructure cost fell to roughly half the initial design, with no practical reliability loss for this workload.
- Microsoft Entra single sign-on, domain-scoped so that only authenticated firm identities are admitted, with Entra group membership driving application-level authorisation.
- Project-scoped retrieval. File search runs automatically within the active project. Web search, where enabled, routes through a provider-neutral API with full-content ingestion and reranking.
Outcome
A platform the firm owns outright across codebase, infrastructure-as-code and deployment pipelines, running across development, staging and production, with all conversation history and uploaded files remaining inside the firm’s Microsoft tenancy.
Stack: React, Node.js, MongoDB, Microsoft Azure (VMs, Key Vault), Bicep, Microsoft Entra, GitHub Actions.