Skip to content

DueCare — Three Deployment Modes

1. Enterprise Integration (Waterfall Detection)

Target: Large platforms (Google, Meta, Discord, job boards)

Architecture: 

User posts / chats / searches
         │
    [Quick Filter]  ← keyword + sentiment (cheap, fast, runs on every message)
         │
    Triggers? ──No──→ pass through
         │
        Yes
         │
    [Gemma 4 E4B Analysis]  ← full DueCare rubric evaluation
         │
    Score < threshold?
         │
    ┌────┴────┐
   Yes       No
    │         │
[Action]   pass through
    │
    ├── Warning popup: "This fee may be illegal under ILO C181"
    ├── Link to resources: POEA, labor ministry, hotline
    ├── Flag for human review (content moderation queue)
    └── Block (if score = worst, e.g., explicit recruitment fee instructions)

Quick Filter triggers (stage 1, runs on everything): - Keywords: "recruitment fee", "placement fee", "salary deduction", "passport held", "training fee", "agency fee", "advance payment", "contract substitution", "kafala", "sponsorship transfer" - Sentiment signals: coercion language, urgency + money, employer-to-worker power dynamics - Context: job postings, DM conversations, group chats about overseas employment

Gemma 4 analysis (stage 2, only on triggered content): - Run through the DueCare guardrails scorer with weighted rubric - Score against all 5 vulnerability categories (BFE, JHE, FCB, PIA, VRV) - Return structured result: score, grade, applicable ILO indicators, jurisdiction-specific law citations

Actions (configurable per platform): - Warning popup with localized legal info ("Under Philippines RA 8042, recruitment agencies cannot charge domestic workers any fees") - Resource links to POEA, POLO, IOM, local labor ministry - Human review queue for content moderation teams - Blocking for worst-grade content (explicit exploitation instructions)

API surface: 

# POST /api/v1/analyze
{
    "text": "I need to pay $3000 to the agency before I can start working...",
    "context": "job_posting",  # or "chat", "search", "comment"
    "language": "en",
    "jurisdiction": "PH_HK"   # optional: origin_destination corridor
}

# Response:
{
    "score": 0.12,
    "grade": "worst",
    "action": "block",
    "indicators": ["illegal_recruitment_fee", "debt_bondage_risk"],
    "legal_refs": ["RA 8042 §6", "ILO C181 Art. 7"],
    "warning_text": "This fee may be illegal. Under Philippine law...",
    "resources": [
        {"name": "POEA Hotline", "number": "1343"},
        {"name": "IOM Migration Health", "url": "..."}
    ]
}

Why multimodal matters for this problem:

Bad actors deliberately use images to evade text-based content filters: - Screenshots of fee structures (bypass keyword detection) - Photos of contracts with exploitative clauses (not searchable text) - QR codes linking to illegal payment portals (opaque to text filters) - Bank transfer receipts showing illegal deductions - Fake agency certificates / forged POEA clearances - WhatsApp screenshots of coercive conversations (image, not text)

Gemma 4's multimodal understanding reads these images. Text-only filters cannot. This is the load-bearing multimodal use case.

Similar to existing systems: - Facebook's "Are you OK?" popup for suicide-risk content - Google's "This search is about a crisis" cards - Discord's age-verification and CSAM detection pipeline - Job board "report this listing" with automated pre-screening

2. Worker-Side Tool (Local/Plugin)

Target: Prospective migrant workers, their families, community organizations

Form factors: - Browser extension (Chrome/Firefox) that scans job postings and chat messages for exploitation indicators - WhatsApp/Telegram bot that workers can forward suspicious messages to for analysis - Mobile app (via LiteRT) that runs entirely on-device — no data leaves the phone - Web app hosted by an NGO (e.g., Polaris, IJM) where workers paste text for analysis

Architecture: 

Worker sees suspicious message / job posting / contract
         │
    [Copy text or screenshot]
         │
    [DueCare Local]  ← runs on phone via LiteRT or via browser extension
         │
    Analysis result:
         │
    ┌────┴────────────────────────────┐
    │  ⚠️  WARNING                    │
    │                                 │
    │  This fee appears to violate    │
    │  Philippine law (RA 8042).      │
    │  Recruitment agencies cannot    │
    │  charge domestic workers any    │
    │  placement fees.                │
    │                                 │
    │  What you can do:               │
    │  • Call POEA at 1343            │
    │  • Contact POLO in [country]    │
    │  • Report to IOM: ...           │
    │                                 │
    │  [Learn More]  [Report This]    │
    └─────────────────────────────────┘

Key design constraint: the worker's text never leaves their device unless they explicitly choose to share it. Gemma 4 E2B (via LiteRT or llama.cpp) runs entirely locally, so the worker can get guidance without uploading raw case content to the public hub.

Multilingual support: Workers from the Philippines, Bangladesh, Nepal, Indonesia, Ethiopia speak different languages. Gemma 4's multilingual capabilities handle Tagalog, Bengali, Nepali, Bahasa, Amharic — not just English.

3. Agency/NGO Dashboard (Custom-Trained Interface)

Target: Recruitment agencies (compliance), NGOs (monitoring), regulators (enforcement)

Architecture: 

┌──────────────────────────────────────────────────────┐
│  DueCare Dashboard                                    │
│                                                       │
│  ┌─────────────┐  ┌─────────────┐  ┌──────────────┐ │
│  │ Batch Eval   │  │ Compliance  │  │ Training     │ │
│  │              │  │ Monitor     │  │ Generator    │ │
│  │ Upload 1000  │  │             │  │              │ │
│  │ job postings │  │ Real-time   │  │ Generate new │ │
│  │ → score all  │  │ feed of     │  │ test cases   │ │
│  │ → export CSV │  │ flagged     │  │ from latest  │ │
│  │              │  │ content     │  │ evasion      │ │
│  │              │  │             │  │ patterns     │ │
│  └─────────────┘  └─────────────┘  └──────────────┘ │
│                                                       │
│  ┌─────────────┐  ┌─────────────┐  ┌──────────────┐ │
│  │ Model Comp.  │  │ Domain      │  │ Reports      │ │
│  │              │  │ Explorer    │  │              │ │
│  │ Compare 5    │  │             │  │ PDF/HTML     │ │
│  │ models on    │  │ Browse all  │  │ compliance   │ │
│  │ same prompts │  │ 74K prompts │  │ reports for  │ │
│  │              │  │ + rubrics   │  │ regulators   │ │
│  └─────────────┘  └─────────────┘  └──────────────┘ │
└──────────────────────────────────────────────────────┘

Features: - Batch evaluation: Upload CSV of job postings / contracts / chat logs → score all against rubric → export results - Compliance monitoring: Real-time feed for recruitment agencies to self-audit their content - Training generator: Create new test cases from the latest evasion patterns discovered by the DueCare agents - Model comparison: Side-by-side evaluation of Gemma 4 vs. GPT vs. Claude on the same prompts (useful for agencies choosing a model) - Domain explorer: Browse the 74K prompt corpus with filtering by category, difficulty, corridor, grade - Compliance reports: Generate PDF/HTML reports for regulators documenting what was tested and how the model performed

Custom training: Agencies can fine-tune Gemma 4 on their specific compliance needs: - A Philippines recruitment agency focuses on POEA regulations - A Gulf state labor ministry focuses on kafala reform compliance - An ILO field office focuses on C181/C189 implementation monitoring

Deployment: FastAPI backend (already in DueCare's architecture) + React/Vue frontend. Can run on-premise or as a hosted service.

How These Map to DueCare Components

Deployment Components Used
Enterprise (waterfall) Quick Filter (new) → DueCare scorer → Action Engine (new)
Worker tool (local) LiteRT/llama.cpp model → DueCare scorer → UI (extension/app)
Agency dashboard Full DueCare stack → FastAPI → Web UI
Component Status
DueCare scorer (weighted rubric) ✅ Built
74K prompt corpus ✅ Extracted
5 evaluation rubrics ✅ Ported
Ollama adapter (local inference) ✅ Built
Pipeline schemas (Pydantic) ✅ Built
Quick Filter (keyword + sentiment) 🔲 Not built
Action Engine (warnings, resources) 🔲 Not built
LiteRT export 🔲 Not built
Browser extension 🔲 Not built
FastAPI dashboard 🔲 Not built (exists in _reference/)
Custom training pipeline 🔲 Not built

Video Impact (70 points)

Each deployment mode is a compelling video segment:

  1. Enterprise: "Imagine if Facebook had flagged this recruitment post before Maria clicked Apply" (5 seconds, mock UI)
  2. Worker: Maria holds her phone, pastes a chat message, sees the warning popup in Tagalog (10 seconds, real demo)
  3. Dashboard: NGO compliance officer runs batch eval, exports report for the labor ministry (5 seconds, real UI)

The three modes together tell the story: from platform-level protection → individual empowerment → institutional accountability.