Who it is for
Kamla ji is in her sixties and takes six tablets a day. When her son forgets to remind her, she forgets, and her doctor sees a gap.
The decisions
| Decision | Because |
|---|---|
| Voice first, on WhatsApp | She uses WhatsApp daily and prefers to speak. No app, no forms. |
| Sakhi calls first | She will not open an app. Sakhi reaches out on its own schedule. |
| One care plan | Diet advice must work for all her conditions at once. |
| A contact budget | At most 3 messages and 1 call a day, with quiet hours. |
Safety is code
In health, one unsafe reply ends trust. So safety does not depend on the model behaving well.
- Message
- checked first
- Red-flag rulesfixed, not a model
- no flag
- Modelno tool can change a dose
- drafts
- Output guardblocks unsafe drafts
- safe
- Reply
- Red-flag rules
- flag
- Fixed protocolapproved text, no model
- sent at once
- Reply
If a check fails, Sakhi does the safe thing. Every clinical text needs a doctor’s approval, tied to the exact text.
In the first runs with a real model, the checks passed all 52 red-team cases with no guardrail violations. The same tests passed while Sakhi still sounded scripted, so I am changing how it is tested.
What is left
- Doctor review of every clinical text.
- Legal review.
- A dry run with real members.
- Tests that measure judgment.