Challenge
Internal tools needed dependable language-model capacity. Third-party APIs meant handing over internal data, accepting variable latency, and paying per token for workloads that ran constantly.
Architecture
Several high-memory GPU servers host a pool of quantized models behind one OpenAI-compatible gateway, so every tool integrates once and never learns about the hardware. A router sends each request to the model suited to it, a queue bounds concurrency per node, and a hosted API stays configured as a fallback.
Technical decisions
Quantization was set per model against measured VRAM headroom rather than applied uniformly, because over-quantizing a reasoning model degrades output in ways that are expensive to debug downstream. Context length was derived from concurrency targets, since both compete for the same memory. A small fast model handles classification, so cheap work never occupies expensive capacity.
Outcome
Internal tools call one endpoint and get whichever model fits the task. Capacity grows by adding a node, and the data stays on infrastructure that is directly controlled.