{03/09} · Clinical software
All workIksha App: Offline Conversational Clinical Registry
Can clinicians talk to a registry without patient data ever leaving the hospital?
Clinicians type or dictate to a local language model (Qwen2.5 or Llama 3.2 through Ollama). Each hospital defines its own schema, and the system generates validation, views and the model's tool set at runtime. It includes AI fundus screening and a literature assistant that cites every statement. Built with Python, FastAPI, React and Tauri.
- 1hospital machine, with no internet
- 0database writes by the language model
- 0hard deletes
What the record supports
Claim
A conversational registry can run on one hospital machine with no internet, so patient data never leaves it.
Evidence
The model's output is grammar constrained, and it can propose changes but can never write to the database. Every value on screen comes from PostgreSQL.
Limits, stated plainly
Writes are versioned, nothing is hard deleted, and an append only, hash chained audit log records each change.