We gave First Technology staff local AI chat and agentic desktop agents running on our private NVIDIA GPU cluster, so the data and token generation stay local, the bill is in local currency, and responses come back fast.
Private AI Inference
FDX - Apex (GPU)
Methodology
Practice
First Technology wanted to put AI in the hands of its own people, a local AI chat assistant and agentic desktop agents that could actually do work on the desktop, not just answer questions. The blocker was the same one everyone hits: the cost of cloud-based inference. Priced per token and billed in US dollars, an organisation-wide rollout to staff would have meant an unpredictable, foreign-currency bill that grew with every conversation. For an interactive assistant, the round trip out to a cloud region also added latency that staff would feel on every request.
We stood the inference up on our own private GPU compute cluster instead of the public cloud. The cluster runs NVIDIA GPUs on the CUDA stack with PyTorch-based serving, hosting an open-source LLM, and the chat and the desktop agents call it directly. Because it is our own infrastructure running locally, the data and every token generated stay inside the environment, and requests do not have to travel out to a hyperscaler and back, which keeps responses quick enough for an interactive assistant and the agents that act on the desktop.
Staff got a local AI chat assistant and desktop agents that feel responsive, because the inference runs close by on our own cluster rather than out in a cloud region. The data and the tokens stay local, the cost comes in local currency instead of dollars, and the low latency makes the assistant something people keep using rather than abandon. First Technology can roll AI out across its people on a cost base it controls.
Similar work in Technology & Innovation
First Technology
A cloud chat assistant, Harvey, that lets legal and paralegal staff check contract terms, dates and notice periods in seconds instead of digging through documents.
Practice
Modernise Apps with Azure AI
Vector Gaming
Vector Gaming needed serious GPU horsepower to train a custom 350-million-parameter model for smarter game NPCs. We put our private GPU cluster behind their Vulkan-based distributed trainer and cut a two-month run down to a fraction of that.
Practice
Distributed Model Training
Urban Connect
Urban Connect’s agentic diagnostic agent works a network fault out for itself, pulling the logs, querying the routers and upstream providers, then writing the call-centre ticket, so support is proactive instead of a manual scramble.
Practice
Agentic Diagnostics
Get in touch
Send us the gist of what you are trying to do. The right person from First Digital will be in touch within a working day.
