Which teams use TRACER in production?

Obside (French fintech, intent-vs-news matching for automated trading, 95% saved vs GPT-5) and getclaw (Hermes-based agent framework, 50% end-to-end agent cost reduction via tool-selection routing). Three more case studies are announcing soon.

What kinds of workloads are case studies for?

Workflows where an LLM repeats the same kind of structured decision at scale: classify, route, tag, match, triage, detect, score, choose. Each TRACER case study is one verb applied to one production workload.

How much LLM cost do TRACER customers typically save?

Reductions span 50% to 95% versus the frontier-LLM baseline, depending on how predictable the workload is. The savings come from routing the predictable slice to a local ML surrogate trained on the customer's own classification traces, while parity with the teacher LLM is gated by measured agreement.

Case studies

Teams who stopped paying frontier prices for repetitive decisions.

Each case is one workflow verb, match, route, classify, choose. The number is real, the partition is auditable per cell, the surrogate ships behind an OpenAI-compatible endpoint. Pick the one that looks like your stack.

Match · Finance · 2026

Obside

One frontier-LLM call per news × intent pair, replaced with a 38-cell TRACER surrogate. 95% saved vs GPT-5. 99.9% routed accuracy. 26,591 labelled pairs, no manual labelling.

Read →

getclaw

Choose · Agentic · in prod at getclaw.sh

getclaw

We rewired Hermes (an open-source agent framework) to route tool selection through a TRACER classifier instead of the LLM. End-to-end agent cost dropped ~50% with no degradation.

Read →

Route · classify · AI infrastructure

Announcing soon

Internal classification for model routing and content-type detection at infra scale. Case study in preparation.

soon

Workflow · TBD

Announcing soon

Repetitive structured-decision workload. Full write-up landing here when the customer is ready to share.

soon

Workflow · TBD

Announcing soon

Repetitive structured-decision workload. Full write-up landing here when the customer is ready to share.

soon