Most companies using AI APIs do not think much about which model handles each request. They pick one, usually OpenAI, point their code at it, and watch the bill grow. Javier Loureiro and Pedro Ansio, two Spanish engineers, spent enough time on the paying end of that arrangement to decide it was worth fixing.
The result is Trovald, a proxy layer that sits between an application and its AI providers. Instead of sending every request to the same model, Trovald evaluates what the request actually needs, picks the cheapest provider capable of handling it well, and returns the response in whatever format the application expects. From the outside, nothing changes. Internally, the economics can look quite different.
Thirty Thousand Requests a Month
Trovald is currently processing around 30,000 requests per month across its beta users, a figure that gives Loureiro and Ansio real data on how routing decisions play out across different task types and providers. The company supports twelve providers in total, including OpenAI, Anthropic, Google, Mistral, DeepSeek, and Groq, and handles the format differences between them internally so that developers do not have to.
The core problem Trovald is addressing is straightforward. A question answered in a customer support chat and a complex code review are both API calls, but they do not require the same model. Routing them identically means paying premium prices for tasks that cheaper models handle just as well. Loureiro and Ansio built a classifier that makes that distinction automatically, without any configuration from the teams using it.
How the Pricing Works
Trovald charges 22% of the savings it generates. If a request would have cost a dollar through a team's existing setup and Trovald routes it for sixty cents, Trovald takes roughly nine cents. If there are no savings on a given request, there is no charge. The model is unusual in the infrastructure space, where most tools charge flat fees or per-seat subscriptions regardless of what they actually deliver.
Billing runs through a prepaid wallet system. Teams deposit credits, which are drawn down per request. Each API response includes a header showing the exact cost of that call, which makes it straightforward to attribute spend to specific features or users within a product.
Reliability and Integration
One concern with routing across multiple providers is what happens when one of them goes down. Trovald includes a circuit breaker that monitors provider health and redirects traffic automatically when error rates or latency exceed normal thresholds. The failover is transparent to the application making the request.
Integration requires changing two things in existing code: the API base URL and the authentication key. Trovald issues keys prefixed with tv_ and accepts requests in the OpenAI format, translating them to whatever format the destination provider requires. Teams that want to test the routing behavior without affecting production traffic can run Trovald in shadow mode, which processes requests in parallel with the existing setup and logs what the routing decisions would have been.
Valuation and Funding
Trovald is now raising its first round of outside capital. The company has set an implied valuation of approximately $3.5 million, a figure derived from the terms of the raise rather than a formal appraisal process, which is standard practice at this stage for early-stage software companies. The funding is intended to cover infrastructure scaling, expansion of the provider network, and the development of a machine learning-based classifier to replace the current rule-based routing logic.
Javier Loureiro and Pedro Ansio have also applied to Y Combinator, the American accelerator whose alumni include Airbnb, Stripe, and Dropbox. Acceptance would provide not only capital but access to a network of potential customers and follow-on investors in the United States market, where the bulk of enterprise AI spending is concentrated.
Data and Privacy
Trovald logs metadata about each request, including timestamp, token count, provider used, and cost. The content of requests and responses is not stored. For teams with stricter requirements around data residency, the full stack is self-hostable, meaning the routing infrastructure can run within a company's own environment rather than passing through Trovald's servers.
