DRW Nexus

The problem

Teams wanted to use large language models. Each provider wanted a separate account, a separate card, and billing in dollars — and the moment more than a handful of people are involved, nobody can answer the two questions that matter: who spent this, and what is the ceiling.

Handing every person a raw provider key solves the access problem and creates a much worse one. There is no cap, no attribution, and no way to turn off one person without rotating everyone.

The approach

One OpenAI-compatible endpoint in front of every provider, built on LiteLLM and run on our own machines. Each person gets a virtual key with a monthly budget attached, so exceeding it stops that key and nobody else.

On top sits the part that makes it usable outside engineering: a rupiah wallet with top-ups, per-user and per-team cost reporting with margin, and a built-in chat client with vision, image and video generation for people who will never touch an API. Data-loss prevention runs on the gateway rather than in each application.

Because everything is one endpoint, adding a model is a configuration change, and every application downstream — including this website''s own drafting pipeline — inherits it without a code change.

The outcome

Running in production as the single model source for every internal application, with usage and cost attributable per person and a hard ceiling that cannot be exceeded by accident.

Adding a new provider no longer means a new account, a new key, or a new invoice to reconcile.