TRACER
TRACER is an open-source framework that trains lightweight surrogate classifiers from an LLM's own production traces, routing predictable classification calls to a near-free model while deferring ambiguous inputs back to the teacher. The system provides calibrated per-region accuracy bounds computed on held-out data, so teams only activate the cheap path when quality targets are met. In one production case, TRACER moved 95% of calls to a small surrogate at parity with the teacher model.
- Artificial Intelligence
- Developer Tools
- Software Only
Funding
Raised to date
$500KRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
Founders
Founder details are not available yet.
Product
Problem
Many production AI systems send every request to a frontier LLM, including high-volume, repetitive classification tasks like customer support triage and firmographic tagging. This creates standing costs that scale with traffic volume, even though most of the work is repeatedly deciding the same small set of intents or categories, and teams lack an evidence-based way to identify which calls are safe to handle with a cheaper model.
Solution
Tracer Research Lab offers TRACER, an open-source framework that trains lightweight surrogate classifiers from a model's own production traces — the inputs and labels the LLM has already produced. An acceptor gate estimates teacher agreement per region of traffic, routing predictable cases to the small classifier at near-zero cost while deferring ambiguous, rare, or out-of-distribution inputs back to the LLM. Each region carries a calibrated lower bound on agreement computed on held-out data, so teams only activate the cheap path when a quality target is met. The hosted version at app.tracerml.ai adds live metering, savings estimates, and one-click activation for certified regions.
Target Audience
Primary users are engineering and RevOps teams at companies running high-volume LLM workloads such as customer support ticketing, firmographic tagging, and other repetitive classification pipelines where inference costs scale with traffic.
Features
- Acceptor gate architecture that learns when a surrogate matches teacher agreement from production traces, with no hand-labelling required
- Calibrated per-region error bounds computed on held-out data, showing dominant label, real examples, and agreement lower bound before routing traffic
- Out-of-distribution gate that automatically defers unfamiliar inputs back to the teacher LLM
- Open-source Python library (`pip install tracer-llm`) that runs the entire flow locally
- Hosted dashboard with live coverage, savings meter, and one-click activation once regions are certified
- Demonstrated outcomes including 83.2% local accuracy on Banking77 with .959 teacher agreement, and 95% cost savings in the Obside case study