DueCare: A Gemma 4 Safety Ecosystem for Migrant-Worker Protection¶
DueCare is open-source software that helps platforms, NGOs, regulators, researchers, developers, and worker-support deployments spot recruitment fraud, illegal fees, coercive control, and trafficking risk earlier. Gemma 4 is the reasoning engine, but the strength comes from the component ecosystem around it: workbenches, harnesses, knowledge packs, search guards, anonymization gates, graph extraction, evaluation, benchmarks, and a public hub that can exchange only safe, reviewed objects.
The ecosystem idea¶
DueCare is not a single chatbot. It is a set of components designed to talk to each other through shared fact objects, evidence edges, versioned packs, traces, and exports. A document review can feed Knowledge Extraction. Knowledge Extraction can feed Search and evaluation. Search results pass through safety and verification before they become context. Anonymization & Sharing is the consent gate before anything leaves a local node.
The strongest use case is a network of trusted local nodes. A worker, NGO office, regulator, platform team, or researcher can process raw case material locally. DueCare helps turn that material into reviewed fact objects, evidence edges, and risk-pattern summaries. Raw worker files stay where they belong. Only explicitly reviewed, redacted, and anonymized facts or aggregate signals move outward.
When many local nodes share those safe fact objects, the ecosystem can see recruitment-abuse and trafficking patterns that no single office can see alone: repeated fee requests, passport-retention clauses, corridor-specific false promises, and cross-case signals that justify stronger prevention, investigation, and worker support.
The Render public hub is the reference website for this ecosystem story:
duecare-ai.com. Its source lives in
apps/duecare-ai.com,
so Pages documentation, Kaggle surfaces, and the public website can be
reviewed against the same repo.
Public URLs¶
| Surface | URL | Role |
|---|---|---|
| Main server website / public hub | duecare-ai.com | Render-hosted FastAPI site and coordination service. It shows the public product story, serves hub APIs, and accepts only public-source proposals, vetted pack metadata, anonymized aggregate signals, hash receipts, and consented contact metadata. |
| GitHub source repo | github.com/TaylorAmarelTech/gemma4_comp | Source of truth for the monorepo: Kaggle kernels, packages, docs, Render website source, scripts, validation gates, and GitHub Actions. |
| GitHub Pages docs | tayloramareltech.github.io/gemma4_comp | Static MkDocs documentation generated from docs/. This is the easiest route for onboarding, install guides, scenarios, architecture, reproducibility notes, and judge-facing review pages. |
Choose Your User Path¶
DueCare uses the same six lanes as the public hub at duecare-ai.com. They are roles in one ecosystem, not separate products: content moderation, case analysis, worker support, research, anonymized knowledge sharing, and custom API implementations all use the same Gemma 4-centered safety stack.
| User Path | What You Need | Where To Start |
|---|---|---|
| I am on a platform safety team | Screen risky recruitment posts, ads, profiles, and messages before workers are harmed. | Enterprise pilot, Deployment guide, and the content screening API. |
| I am an NGO, caseworker, regulator, or legal-aid reviewer | Turn messages, documents, and public rules into grounded drafts, referrals, complaint materials, and corridor updates. | Caseworker workflow, NGO office deployment, and Regulator pattern analysis. |
| I am a migrant worker or prospective worker | Understand suspicious offers, contracts, recruiter messages, fee demands, document retention, threats, and next steps privately. | Worker self-help and the DueCare Journey Android releases. |
| I am a researcher, evaluator, or auditor | Reproduce model behavior, compare safety interventions, inspect pack hashes, and audit every claim from source artifacts. | Researcher analysis, Reproducibility, and A-00 proof path. |
| I am sharing anonymized knowledge | Convert reviewed local evidence into sanitized fact objects that improve shared packs without centralizing raw worker case data. | Anonymization policy, Submission labeling policy, the workbench Anonymization & Sharing page, and the civil-society outreach loop (contribute by replying to an email). |
| I am a developer or integration partner | Embed DueCare into moderation tools, NGO systems, mobile clients, dashboards, or custom workflows. | Install guide, Embedding guide, OpenAPI spec, and Client connect. |
Deployment And Use Cases¶
The same components can be assembled for different operating contexts. Each deployment keeps the sensitive-data boundary explicit.
| Use Case | Local Components | Shared Components |
|---|---|---|
| Platform safety deployment | Content screening API, GREP rules, corridor packs, risk explanations, queue-ready review payloads, and moderation audit traces. | Anonymized pattern signals and vetted pack updates. |
| NGO & regulator office deployment | Local case-bundle review, citation-backed summaries, referral drafts, graph extraction, document review, and caseworker-controlled exports. | Reviewed public facts, sanitized observations, pack proposals, and aggregate corridor trends. |
| Individual worker / mobile deployment | Trusted mobile or chat experience, private notes, local guidance, evidence organization, and worker-controlled sharing. | Only worker-approved notes, intake drafts, or sanitized signals. |
| Researcher deployment | Reproducible prompts, model comparisons, rule and LLM judging, optional community benchmark runs, and A-00 training/evaluation artifacts. | Versioned scorecards, model cards, evaluation metadata, and pack hashes. |
| Anonymized knowledge sharing deployment | Local redaction, PII checks, consent-aware metadata, fact-object review, evidence-edge review, and hash receipts. | Sanitized fact objects, aggregate signals, vetted pack proposals, benchmark rows, and rejected-submission receipts. |
| Developer / integration deployment | FastAPI routes, package modules, schemas, Docker/runtime examples, client snippets, and local validation scripts. | Public APIs, docs, pack registry, release artifacts, and integration examples. |
Information Sharing Architecture¶
DueCare is designed for many local nodes to learn from each other without becoming one raw case-data warehouse.
| Stage | Component | Output | Boundary |
|---|---|---|---|
| Local review | Workbench, mobile app, or tenant deployment | Drafts, citations, graph edges, review decisions, and local audit traces. | Raw chats, case files, IDs, screenshots, and private documents stay local. |
| Local anonymization | Anonymization & Sharing workflow | Sanitized fact object, aggregate signal, public-source proposal, or hash receipt. | Reviewer consent and redaction happen before upload. |
| Hub intake | duecare-ai.com public hub | Review queue entry, validation result, and submission receipt. | Server-side PII checks reject raw private case content. |
| Curation | Curator and civil-society review | Vetted pack update, contact metadata update, benchmark row, or rejected proposal. | Human review is required before shared knowledge becomes reusable. |
| Reuse | Local nodes pulling updated packs | Better rules, citations, examples, and evaluation artifacts in future local deployments. | Shared intelligence returns as versioned packs, not as exposed case narratives. |
Active outreach. The hub does not only wait for submissions — it runs a civil-society outreach loop: it detects which knowledge needs field verification (a corridor fee cap, an emerging payment rail, a statute change), emails a specific question to the opted-in experts best placed to answer, vets each reply through the same PII gate, and folds the observations back into prioritized context and candidate grading dimensions. Civil society contributes by replying to one email — never by learning a new login. Raw addresses are never stored; only a sha256 hash plus topics persist.
Read the focused Information Sharing Architecture page for the object types, trust boundary, and hub/local responsibilities. For a shorter public explanation, read Gemma 4 for Networked Knowledge Sharing Without Centralizing PII. For copy-ready Kaggle discussion copy with a title and opening paragraph, use Copy-ready Kaggle Post: DueCare Networked Knowledge Sharing.
Source verification¶
Recruitment-fraud drafts are only as trustworthy as the entities behind them, so DueCare grounds them in real public records. The entity-intelligence pipeline verifies recruiters, employers, and their beneficial owners against 34 official government registries — licensed-recruiter lists, fishing-vessel and money-lender registers, sanctions and debarment lists, and GLEIF / OpenOwnership corporate ownership — drawn from a 1,111-source catalogue spanning 95 countries and 18 industries, alongside 532 migrant worker-support organisations (helplines, shelters, legal aid, unions) for referral. It is propose-only: resolved records are staged for curator review, never auto-merged, and kept strictly separate from the trafficking knowledge layer.
Judges can browse the live registry map, licence ledger, and connector status at duecare-ai.com/source-verification; the full per-source map and connector reference live in the Entity-intelligence pipeline doc.
Core Components¶
| Component | Purpose |
|---|---|
| Gemma 4 model layer | Local reasoning, classification, summarization, multimodal reading, and draft generation. |
| Safety guidance layer | Persona, GREP rules, retrieval, tools, imports, search verification, and traceable response assembly. |
| Knowledge packs | Versioned rules, citations, contacts, corridor facts, tool definitions, examples, and evaluation data. |
| Local anonymization module | Converts reviewed sensitive material into redacted fact objects or aggregate signals before sharing. |
| Central knowledge server | Hosts public docs, APIs, vetted pack metadata, anonymized aggregate signals, and review queues. |
| Quality and training layer | Reproducible evaluation, A-00 proof runs, benchmark artifacts, and optional Gemma 4 adaptation. |
| Entity-intelligence pipeline | Propose-only verification of recruiters, employers, and their owners against 34 official government registries; feeds curator-reviewed pack updates. See Entity-intelligence pipeline and the live source-verification page. |
Run Or Verify¶
| Goal | Start |
|---|---|
| Try the workbench | Open DueCare App on Kaggle, run it, and open the Cloudflare URL. |
| Review the proof path | Use For Judges and Reviewers and the active A-00 proof path. |
| Install locally | Use Install and Local deployment. |
| Embed in another system | Use Embedding guide, OpenAPI spec, and Deployment enterprise. |
| Audit claims | Use Reproducibility, Notebook guide, and the validation commands linked there. |
Sensitive Data Boundary¶
- Raw worker chats, case files, IDs, contact details, and private documents stay on the worker device, Kaggle session, trusted NGO machine, or tenant-owned deployment unless an authorized user creates a sanitized object.
- Sensitive PII is anonymized by the local workflow before anything is submitted to the public hub.
- The public server receives public-source facts, vetted pack metadata, anonymized aggregate signals, hash receipts, and consented contact metadata.
- The public server should not receive raw worker chats, phone numbers, addresses, passports, private documents, or personal case narratives.
Read the threat model and submission labeling policy for the detailed trust boundary.
Built for the Google Gemma 4 Good Hackathon (Safety & Trust track), submitted 2026-05-18.