DueCare has three jobs: prevent exploitation before it spreads, assist victims and at-risk workers, and help stakeholders understand what is happening and why. Pick the lane that matches your role to see the setup path, docs, and first smoke check.
Prevention lane: screen recruitment and platform content before harm spreads, then hand explained risk signals to your existing review workflow.
Use the classifier and rule packs before harmful recruitment content reaches workers. Your moderation queue decides what action to take.
Use the repo lifecycle target for the chat and classifier services.
Use the platform safety API guide for the structured risk envelope.
DueCare flags and explains; your existing trust-and-safety workflow acts.
git clone https://github.com/TaylorAmarelTech/gemma4_comp cd gemma4_comp make demo make doctor # API shape: POST /api/classifier/evaluate # See docs/deployment_enterprise.md before wiring production traffic.
Run DueCare on an office edge box for case analysis, triage, document bundles, referrals, and regulator pattern review.
The edge example includes Docker Compose, mDNS, backup, and caseworker onboarding.
Caseworkers use office browsers or Android phones connected to the office model endpoint.
Use doctor and backup targets before pilots with real case teams.
git clone https://github.com/TaylorAmarelTech/gemma4_comp cd gemma4_comp/examples/deployment/ngo-office-edge make setup make doctor # Caseworkers open the office endpoint: http://duecare.local
Install DueCare Journey for plain-language guidance, journaling, evidence organization, and reports the worker chooses whether to share.
Download the latest Android release from the sibling app repository.
The app asks for migration stage and corridor, then uses the configured local, on-device, or partner-controlled model path.
Journal, advice, and reports are designed for worker-controlled sharing with trusted helpers.
# Open this on the Android phone: https://github.com/TaylorAmarelTech/duecare-journey-android/releases # Current documented APK: duecare-journey-v0.9.0-twenty-corridors-new-rules.apk
Use the Kaggle notebooks, reproducibility docs, pack hashes, and validator scripts to study patterns and verify the submission.
Judges start with the omni harness chat, then the focused live demo.
Use the repo validator for notebook and evidence-surface checks.
Reports and notebooks should cite model, git SHA, and dataset/version inputs.
git clone https://github.com/TaylorAmarelTech/gemma4_comp cd gemma4_comp make test python scripts/validate_notebooks.py
Pull verified packs, pin versions, submit public-source improvements, or embed DueCare in your own app or workflow.
Start with the live pack registry endpoint, then fetch a pinned version.
Choose Docker, package install, or source checkout depending on integration depth.
Use client submission endpoints for vetted public facts, not private case content.
curl -fsSL "https://duecare-ai.com/api/hub/packs?latest_only=true" \ -H "accept: application/json" curl -fsSL https://duecare-ai.com/api/hub/packs/phl-kwt-domestic \ -H "accept: application/json" # See /client-connect for POST /api/hub/client/submission.
Export only reviewed, sanitized knowledge objects from local evidence graphs. The public hub stores consent-aware facts, hash receipts, and pack metadata, not raw narratives or private documents.
Start from confirmed evidence rows inside the caseworker or partner deployment.
Remove PII, check re-identification risk, and keep consent metadata explicit.
The hub runs a second PII scan before storage and curator review.
curl -fsSL "https://duecare-ai.com/api/hub/signals" \ -H "content-type: application/json" \ -d '{"source":"partner_review","corridor":"PH-HK domestic work","risk_tags":["fee_request"],"summary":"Sanitized aggregate pattern only.","consent_basis":"authorized_sanitized_submission"}'
| Item | Origin | Public hub | Why |
|---|---|---|---|
| Worker message text | Worker / caseworker device | Never crosses | Hub schema rejects free-text case fields. |
| Names · IDs · contact info | Worker / caseworker device | Never crosses | No PII column exists on any hub table. |
| Employer name / contract photo | Worker / caseworker device | Never crosses | Same as above. Identification by edge filter. |
Anonymized pattern_id (e.g. fee_request) | Local model + grep | May cross. Opt-in only | Aggregated only. K-anon ≥ 30 floor. |
| Corridor + sector | Local config | May cross. Opt-in only | Anonymized buckets only. Used for trend rollups. |
| Pack version + pull-time | Local config | May cross | Audit reproducibility for curators. |
| Public-source URLs | You | Crosses (intended) | Public sources. Reviewed by curator. |