Cases
One model on hardware the client owns, one in the client’s own AWS account.
SSN · On-premise
SSN sells data-center servers, storage and networking, and none of their data was permitted to leave their building. We fine-tuned an open-weight model and serve it on a fixed set of GPUs on their own hardware: over 5 billion tokens a month at over 90% lower spend than frontier APIs. On it runs an AI configuration agent with a local knowledge base, kept current by a pipeline that syncs vendor documentation. It works out which parts fit together, checks live ERP inventory and prices the result. At five minutes a quote instead of two hours, small deals are now worth fulfilling. SSN runs over 800 a month. A second service tracks competitor listings and adjusts pricing, previously a daily manual task.
Tactical IC · AWS private cloud
Tactical IC sells integrated circuit components to manufacturers and labs. They wanted AI in their sourcing without sending data to a third party or paying per token at scale. We fine-tuned the latest Kimi model and deployed it on autoscaling instances in their AWS account. On it runs a sourcing agent that watches the sales and support inbox, checks each request against their private database of vendors and price points and against public catalogs, and hands back a complete purchase order request for approval.
Tell us what you send to hosted AI today and roughly what you spend. We’ll tell you what can move in-house and what it would cost to run.