Longhand by Tech Leaders
Owned AI deployment guide
Owned AI is Longhand’s deployment option for training or fine-tuning private models and running inference on hardware your organization controls. Use those models in Longhand workflows with review steps your team configures. You can choose cloud models for other work.
Run the classification exampleWhere the data goes
- Training and fine-tuning
- Select examples your organization has permission to use. The training job processes those examples in the environment your organization controls.
- Private inference
- A workflow step sends its prompt and context to the selected model endpoint. Select a compatible model running on your hardware for steps that need private inference.
- Cloud models and connected services
- A cloud model can receive task context from a workflow step. Trial and company AI requests can pass through Longhand’s servers. Connections such as Google Workspace still contact their providers when the model runs locally.
The Google API Limited Use policy prohibits using Google Workspace API data, including derived data, to develop, improve, or train generalized or non-personalized models. Check provider terms and settings for the data your workflow uses.
What to prepare
- One workflow and a labeled sample. Pick a repeatable task, such as classifying records. Collect approved examples, expected labels, and ambiguous cases that should go to review.
- A model and a compatible runtime. Longhand desktop supports its bundled
llama-serverand OpenAI-compatible local endpoints for inference. Confirm that the model supports the tools your steps use. - Hardware sized for your workload. Scope training and inference separately. Model size, data volume, and expected concurrent use determine what you need.
- A data map. Identify where training files, prompts, connector results, outputs, and exports will be stored or sent. Review model fallback and tool trust settings before using sensitive inputs.
How to verify a pilot
- Set a baseline. Run approved samples on an existing compatible model. Compare labels with expected results and record review time. If you fine-tune, hold back examples from training and use them for evaluation.
- Check routing. Assign the private model to the steps that need it. Confirm the model in Longhand’s run history and check private endpoint logs where available. Test what happens when the endpoint is unavailable. A local-preferred setting can fall back to a cloud model.
- Exercise human review. Configure approval before uncertain or consequential output moves forward. Run an ambiguous case and verify that the workflow pauses for review. Try both approval and rejection before connecting a live destination.
- Check the export. Export the workflow and record the model endpoints and connection permissions needed to run it. Review the exported files for sensitive data before sharing.
Try reviewing a synthetic support request and downloading the decision in the runnable classification example. The example classifies requests in your browser with simple rules.