Turn a customer email into a safe CRM update.

Choose a fictional message. A real local model reads the email, the customer record, and the handling rule, then proposes the exact status and queue changes. Nothing is written until you approve the preview.

Run the local AI demoOpenAPI documentInterface referenceMCP endpointRead the teardown
Runs locally with fictional dataCustomer Request Lab
Preparing the demo
Customer inbox

Choose an email

Each message maps to a real back-office decision.

The demo inbox is empty. Reset the demo to reload it.

Customer message

Choose an email

Select an email to open its customer record and handling rule.

CRM write preview

Review the exact update

The model can suggest a change. Only this interface can preview and apply one.

Run the local model or use the guided route. The before-and-after record will appear here.

Local AI Demo

Give a real local model a customer email. See the CRM update it proposes, the rule it checks, and the exact change before anything is written.

Load the model

Qwen2.5-0.5B-Instruct runs in this browser through WebGPU. It reads only the fictional email, CRM facts, and handling rule. No prompt is sent to an external model API.

Download
about 520 MB, once per browser cache
License
Apache 2.0
Data
Fictional customer email and CRM record
Permadyn configured the task harness, policy boundary, and interface. The underlying model was not trained by Permadyn. Desktop Chrome or Edge with modern graphics support is recommended.

The demo is opening a synthetic case. You can write an instruction while it finishes.

Suggested CRM update

The result will show the proposed queue and status, the facts used, and the handling rules checked.

Observable evidence

What the interface recorded

The model reads the message. The application owns the update.

The local model receives only the fictional email, CRM facts, handling rule, and visitor instruction. It never receives a credential or a function that can change state. Validation, visible preview, confirmation, version checks, and idempotency remain separate controls.

  • The model runs locally with no external inference API
  • Customer emails and model output are treated as untrusted input
  • The result is limited to known CRM fields and queues
  • Every write is rule checked, version bound, and idempotent
  • No prompt or model output is sent to analytics