Skip to content

Duecare Channels — NGO / Government chatbot integrations

Status: Roadmap. Design target. Today's bridge: packages/duecare-llm-chat already exposes the harness through a FastAPI surface — that's the integration surface Channels would wrap with platform adapters. No code in this repo currently sends messages on Messenger / WhatsApp / SMS.

A channel-deployment layer that lets trusted institutions (NGOs, labor ministries, consulates, regulators, worker centers, hotline operators, legal-aid orgs) provide Gemma 4-powered migrant-worker guidance through familiar messaging platforms — grounded in their own verified RAG documents, contacts, complaint mechanisms, and intake workflows.

Why Channels is its own component

The mobile app is worker-owned. Channels is institution-owned.

Module Owner User Deployment surface
Mobile (component #8) Duecare / worker individual migrant worker phone (private, on-device)
Channels (component #7) NGO / government tenant migrant workers + caseworkers Messenger, WhatsApp, web chat, SMS, embassy portal

Both use the same safety engine (Runtime + Harness + Eval). The trust model differs: a worker downloading the Mobile app trusts Duecare directly. A worker chatting with an NGO bot trusts the NGO, which has wrapped Duecare. Channels ships with explicit "who-you're-talking-to" disclosure on session start.

Target channels (roadmap, in order of partner demand)

Channel Notes
Web chat widget Embeddable on NGO + government websites; lowest friction
WhatsApp Business Highest-reach for migrant-worker populations; needs Meta Business approval
Facebook Messenger Used by recruiters too — caseworker triage usable here
Telegram Strong in Eastern European / Russian-speaking corridors
Viber Strong in Philippines + diaspora communities
SMS (Twilio / Africa's Talking) Last-mile reach; works on any feature phone
Government website widget Embassy / consulate / labor-ministry intake portals
NGO intake page IFRC / IOM / IJM / Polaris / etc. existing portals

What the chatbot does

Function Description
Worker Q&A Answers worker questions in plain language
Intake triage Collects structured complaint facts
RAG grounding Uses NGO / government-specific documents
Contact routing Shows verified agency / NGO contacts
Complaint guidance Explains how to file safely
Draft generation Creates complaint / referral drafts
Document review Reviews screenshots, receipts, contracts
Escalation Hands off to a human caseworker
Safety planning Advises against unsafe confrontation
Multilingual support Tagalog, Bahasa, Nepali, Bengali, Arabic, etc.

Tenant knowledge pack (per-deployment)

Each NGO / government deployment has a tenant-specific knowledge pack:

NGO / government chatbot tenant
│
├─ official FAQ
├─ complaint forms
├─ intake scripts
├─ jurisdiction laws
├─ contact directory (subset of _contacts.json)
├─ escalation rules (caseworker handoff thresholds)
├─ supported languages
├─ agency-specific disclaimers
├─ human handoff contacts
└─ reporting thresholds (when to flag for caseworker review)

Combined with the global Duecare knowledge (ILO conventions, Palermo Protocol, common indicators) plus the continuously updated intelligence delivered by the Public Information Research Monitor through vetted Exchange packs.

Complaint-mechanism boundary

Allowed

  • Explain complaint options
  • Identify likely agency / NGO contacts
  • Prepare a draft complaint
  • Pre-fill a form (only when the form supports URL params)
  • Hand off to human staff
  • Let user choose whether to submit
  • Save a case ID after explicit consent

Not allowed by default

  • Auto-submit complaints
  • Auto-email regulators
  • Auto-call hotlines
  • Expose raw messages to a central server without consent
  • Tell workers to confront employer / recruiter
  • Hallucinate contact info
  • Pretend to provide legal advice

Canonical safety phrase, used in every channel's session-start banner:

Duecare drafts; the user or trusted caseworker decides. Raw chats stay with the tenant unless explicit consent allows a handoff.

Architecture sketch

Worker on Messenger / WhatsApp / web / SMS
        │
        ▼
   Channel Adapter        (per-channel webhook + session manager)
        │
        ▼
   Tenant Router          (loads tenant knowledge pack)
        │
        ▼
   Consent UX             (session-start: who you're talking to + safety)
        │
        ▼
   Duecare Harness        (GREP + RAG + tools — same as the chat package)
        │
        ▼
   Gemma 4 Runtime        (per-tenant model: stock or tuned via Trainer)
        │
        ▼
   Response policy        (grounded refusal, contact routing, draft gen)
        │
        ▼
   Channel Renderer       (markdown → channel-specific formatting)
        │
        ▼
   Worker reply           (or human handoff if escalation threshold met)

What Channels does NOT own

  • Model training (that's Trainer / component #5)
  • Global laws database (that's RAG via Harness / Exchange)
  • Rubrics (that's Eval / component #4)
  • Raw source crawling (that's the Research Monitor / component #6)
  • Official contact verification (the Research Monitor proposes; humans verify)

Channels' job is session management + tenant config + channel adapters + safety UX. Everything that makes a response correct lives upstream.

Submission disclosure

The Kaggle submission does not ship Channels. Auto-deploying Duecare on Messenger / WhatsApp without partner approval would be unsafe — it would create the impression that the bot is an official NGO channel when it isn't. Channels exists in this doc tree as a design target that partners can review.

A live Channels deployment requires:

  1. A named NGO / government tenant who owns the bot.
  2. A tenant knowledge pack (FAQ + complaint forms + contacts + handoff).
  3. Platform business-account approval (Meta Business / Twilio / etc.).
  4. Eval pass + vetted pack pinning.
  5. Disclosed safety boundaries in the session-start banner.

None of those happen at hackathon time.