Zinroute

LLM bills grow quietly and in the wrong places. Most of the spend is not the hard requests, it is the easy ones being sent to an expensive model because that is what the code was written against a year ago. The obvious fixes are all bad: rewriting call sites for every provider, maintaining routing logic inside the product, or accepting a bill nobody can attribute to a feature. The real constraint is that any solution has to be adoptable in one line, because a cost tool that requires a refactor never gets installed.

• Implemented the OpenAI wire protocol exactly — chat, responses and embeddings — so adoption is a base-URL change. The existing SDK, request shapes and response handling all keep working, which is what makes the tool installable on a busy afternoon.
• Put routing behind that stable interface: requests are matched to an appropriate model per policy, reaching many providers through OpenRouter, so the choice of model becomes an operational decision rather than a code change.
• Made spend attributable by design. 11 data models carry project and usage structure, so cost is reported per project and per feature instead of arriving as one opaque monthly number.
• Built waste detection as a first-class capability rather than a dashboard afterthought — surfacing the requests that were over-served by an expensive model, which is where the recoverable money actually is.
• Added a caching layer in front of repeated work, since the cheapest LLM call is the one never made.
• Shipped the full SaaS spine around it: 10 auth endpoints, project and member management, billing, payment processing and webhooks, with guard-protected routes and audit logging throughout. A cost tool holds API keys, so access control was structural rather than added later.

• A live proxy adopted by changing one base URL, with no application code rewritten.
• 51 endpoints across 25 modules covering routing, analytics, billing, payments and team management.
• Per-project cost and usage visibility, so LLM spend can be attributed to the feature that caused it.
• Model routing as an operational lever, letting spend be tuned without shipping a release.