Build the intelligence your work needs.

Use the right model for the job.

Bring in a strong general model for a complex question, use an open-weight model through an approved route, or consult a specialist with a focused task.

Geyser keeps the Agent’s working identity and context separate from the model. Supported models and runtimes can change while the team keeps its files, history, and responsibilities.

• Hosted models through approved provider routes.
• Open-weight and local inference on supported services.
• Specialists available to the agents and jobs you choose.

Begin with a job you can explain.

“Sort these requests using our categories.” “Extract these fields from our documents.” A useful specialist starts with a narrow task and a clear definition of a good answer.

Describe that task in conversation. Agents can help prepare examples, review the data, set aside a test set, and organize the training work. You can guide the process without writing a training pipeline.

Find out whether it learned the right thing.

A candidate is compared with a baseline using held-out examples. Review disagreement, poor results, and the cases that matter to the job. Keep the accepted version connected to the data and evaluation that produced it.

Make the specialist available through an approved route, follow its results, and return to a previous version when needed. Training, evaluation, deployment, and rollback belong to the same workflow.

Run the intelligence where it fits.

Private training supports eligible Apple silicon MLX and Linux LoRA workflows. Geyser also supports separately configured external training and inference services. Model hosting lets the team use an accepted model after training.

The training workflow is in private preview. The model, hardware, license, and intended use determine the available path. A private endpoint hosted by an outside provider is still an outside processing service.