Chatbots & AI agents

AI assistants over live operational data, served on private LLM infrastructure through governed MCP interfaces — every tool call authenticated, authorized and audited, and no tokens leaving the building.

MCP

operational data and actions published as governed tools any approved agent can call

On-prem

inference on private GPU infrastructure — customer data never leaves the perimeter

Governed

every tool call authenticated, authorized and audited

What it does

BitSwan exposes operational data and actions as MCP servers — a governed tool interface that AI agents can call. The first assistant taking shape on top of it:

• TESU ops assistant (in build) — chat over incidents, KPIs, alarms and customer impact for the NOC and technical support, with guided automations that always keep a human approval in the loop and an audit trail behind every action.

The same interface is open beyond internal tools: partner-built assistants can consume the operator’s governed MCP services without ever touching raw systems. All inference runs on private LLM infrastructure inside the operator’s perimeter.

At a glance

Status: Live

Deployed at: CETIN SK

Domain: Customer & Ops

Built with: Private LLM · MCP servers · BitSwan agents · Keycloak / AD

Why it matters

The blocker for telco AI agents was never the model — it was giving the model safe hands. Subscriber data cannot flow to a public API, and an agent that can touch the network needs the same identity, authorization and audit discipline as a human operator. The MCP layer provides exactly that: each tool an agent can use is explicitly published, access-controlled through Keycloak/AD, and every call is logged.

The architecture also decouples who builds the assistant from who owns the data: third-party applications consume operator-governed MCP services — a pattern that scales from internal tools to partner and B2B scenarios.

How it works

BitSwan pipelines keep the operational datasets fresh; MCP servers publish curated views and actions; the private LLM platform (see the Private AI vision) hosts the models; agents — BitSwan’s own or third-party — compose the two. Guardrails, approval steps for any state-changing action, and full audit trails are part of the serving layer, not an afterthought.

Why it replicates

The MCP pattern is the replication mechanism itself: once your data is on the platform, every new assistant — internal or partner-built — is configuration on top of the same governed interface rather than a new integration project.