The language layer
Translates every message. Never touches a dose.
Human-approved · flag-gated per tenantA real day with it
A discharge instruction goes out at 6 PM.
Kannada-speaking Lakshmamma receives it in Kannada — with "1 tablet twice daily after food" numerically untouched, guaranteed by code, and auditable at /admin/sahayak-renders.
- 1
Any outbound patient message is about to send, from any module.
- 2
The agent renders it into the patient's language inside a 2500 ms budget.
- 3
VerbatimGuard protects doses, numbers and links; 5% of renders get QA'd.
- 4
The patient reads the message in their own language, numbers untouched.
Admin
One toggle covers every notification channel.
Patient
Instructions in their own language.
Compliance
Every render logged with QA sampling.
- Flag
- aiagents.sahayak.enabled
- Reads
- The outbound patient message, immediately before it sends
- Never writes
- Never touches a dose, a number or a link — OTPs are never translated
- Channels
- Every outbound patient notification channel
- Languages
- Hindi, Marathi and Kannada; English is the implicit fallback
Questions about this agent
Can Sahayak mistranslate a dose or a number?
No. VerbatimGuard protects doses, numbers and links by code, and OTPs are never translated at all.
How is quality checked?
5% of renders go through back-translation QA, logged and auditable at /admin/sahayak-renders.
Which languages does it cover?
Hindi, Marathi and Kannada today, with English as the implicit fallback rather than a guess.