Abhibhaw Asthana

Building now / Consumer health · Voice AI / In development

Sakhi: an AI nurse that calls first

Older Indians with diabetes, BP and thyroid miss medicines and readings unless family reminds them. Sakhi is a voice-first AI nurse that calls them on WhatsApp in Hindi.

User
“Kamla ji”: 45 to 72, Hindi first, reads a little and prefers speaking. Two to four chronic conditions, 3 to 8 tablets a day. Uses WhatsApp. A son or daughter is the caregiver.
Problem
Readings and medicines get logged only when family nudges. Apps and forms do not work for her.
Decision
Proactive voice calls on WhatsApp. One care plan across all her conditions. Deterministic safety checks run before and after the model on every message.
Not built
Diagnosis, prescribing or dose changes, ever. No app to install, no public signup, no clinician dashboard yet.
Status
In development. Not live with members. Doctor and legal review are open.

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

DecisionBecause
Voice first, on WhatsAppShe uses WhatsApp daily and prefers to speak. No app, no forms.
Sakhi calls firstShe will not open an app. Sakhi reaches out on its own schedule.
One care planDiet advice must work for all her conditions at once.
A contact budgetAt 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.

Every message, call turn and reading
  1. Message
  2. checked first
  3. Red-flag rulesfixed, not a model
  4. no flag
  5. Modelno tool can change a dose
  6. drafts
  7. Output guardblocks unsafe drafts
  8. safe
  9. Reply
When a rule fires
  1. Red-flag rules
  2. flag
  3. Fixed protocolapproved text, no model
  4. sent at once
  5. Reply
Safety runs as code around the model. A red flag never reaches the model.

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.