Agentic automation

Agents that classify, route, call tools and decide across your workflows, with minimal human interaction.

68%

Lower inference cost. Knowunity, at hundreds of millions of requests a month.

81% → 93%

Accuracy. Against the Gemini 2.5 Flash Lite model it replaced.

Read the Knowunity case study

Narrow tasks, repeated at very high volume

Each step of an agentic workflow is a narrow task: label the request, pick the tool, decide what happens next. A frontier API charges frontier prices for every one of them.

A small model trained for one task matches or beats the large one on that task, runs on hardware you choose, and costs a fraction per request.

Where it applies

Request classification and routing

Label every incoming request so it reaches the right workflow. Knowunity runs this at over 100M requests a month.

Tool and API calls

Emit the structured call that looks up an order or runs a query. Rocketgraph generates graph queries on-premises on IBM Power.

Decision support

Pre-assess cases against your policy and hand the hard ones to a person. Octodet classifies security logs at 88% accuracy, against 63% for the LLM baseline.

Pick one workflow to start

Bring the agent that runs it. We build the data and train the first model.