AI & Agents
Production AI is an engineering job. We build the harness that lets agents act safely.
At a glance
Production AI is an engineering job. Agents that act need a harness: knowledge, tools, guardrails, evaluation, routing, observability and a hand-over to people. NGT builds and runs it on hyperscaler, self-hosted or in-region models; engineers sign off every release.
Knowledge, MCP tools, guardrails, routing, evaluation and hand-over are designed in from the first sprint.
Specialised agents behind one super-agent take intake, data queries, monitoring and change work, each through governed tools.
AI speeds up our delivery. Architects own the design, and every change passes review and quality gates.
What an agent needs around it
A reference multi-agent harness for IT operations. Select a component to see what it does.
- KnowledgeHybrid retrieval over a vector store and a knowledge graph, fed from the source systems.
- Tools through MCPAgents act only through an MCP gateway, with a policy check on every call.
- Guardrails and routingChecks on input and output; one model router with a budget per agent.
- Hand-over and auditLow confidence or a sensitive step goes to a person. Every call is logged.
Point at a component to light its links; select it for what it does here and the practice that makes it right on AWS.
Problems we solve
Six problems customers bring us, each with the architecture we use. Select one.
Answers buried in documents and an old chatbot
Staff spend too long searching internal documents. The old chatbot gives loose answers, and any replacement must plug into the systems already in place.
- Documents fetched from the systems that own them, kept current
- A relevance agent checks scope; answers cite their sources
- Wrong answers become test questions a release must pass
Answers that stay in scope and cite their sources, without disrupting the systems around them.
Point at a component to light its links; select it for what it does here and the practice that makes it right on AWS.
Explore AI & Agents
6 pages, each a few minutes to read.
Frequently asked questions
What customers ask first.
Both, but the production work is agents that act: they log and route requests, run data queries, watch services and guide failure handling.
A knowledge assistant over your documents is often the first step, and the same harness carries over.
Whatever fits your data rules: Azure OpenAI, Amazon Bedrock, Google Vertex AI, open-weight models on GPU Kubernetes, or in-region endpoints in Vietnam and Japan.
One LLM gateway sits in front of them, so routing, budgets and audit work the same way for every model.
Guardrails on input and output, a policy check on every tool call, tools only through an MCP gateway with single sign-on, evaluation before every release and a hand-over to people.
Sensitive actions wait for a person, and every call is logged.
Yes. Two delivery patterns: no customer data offshore, where only reviewed packages cross into your zone; and one agent, two environments, where production runs on a self-hosted open-weight model in your network.
With a PoC on one workflow and a success measure agreed up front, before any wider rollout.
Every later release must pass the same test set without regressions.
No. They are run by FPT Smart Cloud, a sister company in the FPT group; CodeVista is built by FPT Software's AI Center. NGT designs, integrates and operates on them when they fit.
