Customer stories

Norlite's underwriting teams were early AI adopters — too early, in a sense. Three departments held three different provider accounts, keys lived in shared vaults, and the monthly reconciliation meeting had become "an archaeology exercise," as their platform lead puts it. Nobody could say what a single quote-summarization actually cost.

They moved every workload behind SwitchGate in a two-week migration that was mostly deleting code: one base URL, keys reissued to named underwriters, a $75 monthly budget per user, and claude-sonnet locked as the only model for anything touching client data. Francophone client correspondence now routes through French-aware auto and reads, in their words, "like it was written in Lévis, not translated in a hurry."

Papineau runs LLM-driven NPC dialogue in a live multiplayer title — thousands of small completions a minute, latency-sensitive, and famously vulnerable to the failure mode every agent team knows: a dialogue loop that retries itself into a four-figure night. Their previous setup discovered these loops on the invoice.

On SwitchGate, the kill switch discovered the next one at 2:13 a.m. and paused the offending key before it spent its fifth dollar. Dialogue traffic runs on fast, cheap models through cross-provider failover (players notice a 900 ms stall; they do not notice a provider swap), while the writers' room drafts lore on frontier models under a separate team budget. Cache hits on repeated barks bill zero.

ClinChart's product summarizes clinical notes — data their compliance team rightly refuses to send to a third-party model API. For a year, "on-prem or nothing" meant "nothing shipped," because self-hosting a model was easy but wiring it into products, budgets and audit trails was not.

The device agent changed the shape of the problem. Their fine-tuned Qwen deployment stays in their VPC on their GPUs; the agent's outbound tunnel makes it callable as private/clinchart/notes-qwen through the same API their teams already use for non-sensitive workloads on cloud models. One catalog, one ledger, two very different data paths — and the audit export finally satisfies both the CISO and the auditors in the same quarter.