Safety engines for WhatsApp health assistants.
Escalation, localization, and content safety patterns for healthcare chat on WhatsApp — from shipping Myna Bolo in production.
·9 min read
TL;DR
WhatsApp health bots need a safety engine, not only a prompt: intent routing, multilingual refusals, escalation to humans, content filters, and eval fixtures for jailbreaks and medical overreach. Brandlabs shipped these patterns on Myna Bolo for women’s SRH support across Hindi, Hinglish, Marathi, Telugu, and English.
Why is WhatsApp a harder safety surface than a web chatbot?
Users treat it like texting a person, messages are short and code-switched, media and voice notes appear, and the UI gives fewer cues that an AI has limits. A polite wrong answer in WhatsApp feels like advice from a friend — which is exactly the risk.
Myna Bolo (with the Myna Saheli assistant) had to work for rural and under-reached users who already live in WhatsApp. That meant safety could not be an English system prompt hoping for the best.
We engineered conversational layers, prompt patterns, n8n/Turn.io workflows, and safety engines so the bot could help without pretending to be a doctor.
What belongs in a WhatsApp health safety engine?
Classifier or rules for emergency and out-of-scope medical asks, locale-aware refusal copy, human escalation paths, rate limits, logging with redaction, and a regression suite that includes vernacular jailbreaks. The model is one component — the engine is the product.
Core loops:
- Detect intent / risk class
- Retrieve only allowed knowledge
- Generate within policy
- Post-check for prohibited claims
- Escalate or refuse cleanly
- Log for review (without leaking PHI into shared tools)
If step 4 is “trust the LLM,” you do not have a safety engine.
How do you keep localization from weakening safety?
Every safety fixture needs translations and code-switch variants. English-only red teams miss the prompts real users send. Update refusal and escalation copy with native speakers, then lock them in CI.
Operational habit: when a new failure appears in Marathi voice notes or Hinglish slang, it becomes a fixture the same week.
Safety debt compounds faster on WhatsApp because volume is high and tone is intimate.
Frequently asked questions
Can we use a generic GPT WhatsApp plugin?
Not for health without a safety architecture, evals, and clear escalation. Plugins optimize connectivity, not clinical trust.
Should the bot ever give dosages?
Default no for consumer SRH assistants unless a governed clinical program explicitly owns that scope. Refusal plus human path is usually the honest product.