UMIS-D: a clinic agent bound to the user's role

Small clinics run on the same chains of clicks all day: find the patient, register them if they're new, book a slot with the right doctor, add the services, take the payment. Each step sits on a different page. UMIS-D is a clinic system (patients, visits, billing, shifts, stock) with an agent built in. A receptionist types "register Julia Sanchez, Thursday 4pm with Dr. Smirnov, consultation, paid in cash" and the agent does it: it searches for her, creates the record, checks the doctor's free slots, books the visit, adds the service and records the payment. The agent never touches the database directly. It can call 11 commands, and every call goes through a gateway that checks whether the agent is paused, whether that command is switched off, whether this user has the permission, whether the arguments are valid, and whether this user may touch this particular visit or patient. A receptionist can't verify a payment through the chat any more than through the payments page. Cancelling a visit and verifying a payment wait for a person to approve the exact arguments. Every step, allowed or denied, goes into the audit log under the user's name. An admin can pause the agent, set its reasoning effort, add instructions, cap steps per run and require approval for any command. These settings can only make it stricter. It runs on one Vultr Cloud Compute VM with Docker Compose: Caddy for HTTPS, Django on Gunicorn, and PostgreSQL. The model is DeepSeek V4.1 Flash on Vultr Serverless Inference, called through the OpenAI-compatible API with tool calling.