Tracer

Teams that deploy AI often default to a single large model for every task, even when many inputs are simple and predictable, which wastes compute and money. Deciding when a classification or routing decision truly needs a full LLM call versus a cheaper local prediction is an unsolved systems problem. Separately, buyers and sellers of proprietary datasets and AI capability lack clear methods to translate technical performance into a defensible price, as shown by Tracer's own analysis of a data-and-software deal where the selected price implied a 0.236% break-even hurdle against conditional value estimates of $50 million to $200 million.

Tracer builds ensembles of specialist open models that work together as a single system rather than relying on one large general model. Echo, its current adaptive AI system, exposes this ensemble through one OpenAI-compatible endpoint and can be customized to a given workload. Alongside the product, Tracer publishes open research: TRACER OSS is an open-source classification-routing project that trains a lightweight local surrogate from a model's own production traces, paired with an acceptor gate that estimates teacher agreement and decides whether to answer locally or defer to the original LLM. Reported results include 83.2% local coverage on Banking77 at .959 teacher agreement, while MNLI was refused in every tested configuration, showing the system's willingness to defer when it cannot meet a confidence target.