Poki SafeTalk

Culturally-aware, crisis-safe AI conversation in English, Pidgin, Yoruba, Igbo, and Hausa — built on Meta's Llama 3.3-70B via Groq, with a guaranteed, deterministic handoff to a verified Nigerian crisis helpline whenever risk language is detected.

This is the same safety-first conversational engine built for Poki, an AI mental health companion for Nigerians — packaged here so other builders (chatbots, HR/wellness platforms, telehealth tools) can integrate safe, native-language emotional support without building the multilingual and safety architecture from scratch.

Why this exists

Most AI infrastructure ignores Nigerian Pidgin and indigenous languages entirely, and most chatbot safety relies on hoping the model itself catches a crisis. This Actor solves both: genuine conversational fluency in five languages, and a crisis-detection layer that sits outside the AI model — so a mental-health emergency is never left to an LLM's judgment alone.

Input

Field	Type	Required	Description
message	string	Yes	The user's message to respond to
language	string	No (default: english)	One of english, pidgin, yoruba, igbo, hausa
conversationHistory	array	No	Prior turns for context: [{ role: 'user'|'assistant', content: string }]. Only the last 10 are used.

Example input:

{
  "message": "I've been feeling really overwhelmed lately",
  "language": "pidgin"
}

Output

{
  "reply": "Ah, I hear you my friend...",
  "language": "pidgin",
  "riskFlag": false,
  "source": "llama-3.3-70b"
}

If riskFlag is true, reply contains verified Nigerian crisis-helpline contact information instead of a normal conversational response — the AI model is bypassed entirely for that turn.

source indicates how the reply was generated: llama-3.3-70b (normal AI response), safety-layer (crisis detected, deterministic response), demo-mode (no API key configured), or error (the AI call failed).

Setup

This Actor requires a GROQ_API_KEY environment variable, set as a secret in the Actor's Settings > Environment variables in Apify Console. Get a free key at console.groq.com.

Optionally set GROQ_MODEL to override the default (llama-3.3-70b-versatile) — check console.groq.com/docs/models for current available models, as Groq's catalog changes over time.

Example use cases

• Add safe, native-language mental health support to an existing HR or student-wellness chatbot
• Build a WhatsApp or Telegram bot for community mental health support in Nigerian languages
• Add a crisis-safety layer to any general-purpose chatbot serving Nigerian users

Known limitations

Crisis-keyword detection currently has direct pattern coverage for English and Pidgin. Yoruba, Igbo, and Hausa crisis phrasing is not yet keyword-matched — the model is still instructed to respond carefully in all five languages, but guaranteed deterministic routing currently only fires on English/Pidgin triggers. Extending this with native-speaker-verified phrases is a priority next step.