Duecare — canonical product definition¶

Duecare is a Gemma 4-powered safety infrastructure platform for migrant-worker protection. It combines local and tuned Gemma models, deterministic safety harnesses, grounded RAG, evaluation rubrics, privacy-preserving knowledge exchange, continuous law / contact intelligence, institution-deployed chatbot channels, and worker-facing mobile assistance so platform safety teams, NGOs, regulators, individual workers, researchers, and integration partners can prevent exploitation, assist victims and at-risk workers, and understand what is happening and why.

Shorter:

Duecare is a private AI safety platform for migrant-worker protection: tuned Gemma 4 models, grounded safety harnesses, reproducible evaluation, continuously updated knowledge, NGO / government workflows, worker-facing mobile assistance, and developer integration paths.

This is not an OFW-specific product. Filipino overseas workers are one important population (and the demo persona for the Kaggle notebook), but the actual user category is migrant workers globally.

Canonical three outcomes and six lanes¶

Lead mission/demo/intro copy with the three outcomes:

Prevent exploitation before it spreads.
Assist victims and at-risk workers.
Understand what is happening and why.

Use this exact lane order everywhere in the repository, website, video, writeup, and notebooks when summarizing Duecare audiences:

Plain-language component naming and communication flow are maintained in canonical_use_cases_and_components.md.

Platform safety — social media companies, job platforms, and marketplaces need scalable detection of exploitative recruitment content. Surface: moderation API, risk scoring, reviewer dashboard, policy audit trace, anonymized pattern signals.
NGO & regulator — NGOs, government bodies, consulates, and labor regulators need case analysis, triage, referral, and grounded summaries. Surface: case-analysis console, contact routing, report drafts, audit/export, vetted knowledge-pack updates.
Individual worker / mobile — migrant workers need private help understanding suspicious offers, messages, documents, and recruiter pressure. Surface: mobile app, private checker, offline support, official contacts, localized rights guidance.
Researcher — academic researchers, evaluators, Kaggle judges, and auditors need reproducible safety benchmarks and model evaluation. Surface: evaluation harness, rubrics, notebooks, model comparison, domain packs.
Anonymized knowledge sharing — reviewers and partners need to convert sanitized, consent-aware facts into reusable knowledge objects without moving raw case files into the public hub. Surface: local anonymization, submission receipts, curator review, and vetted pack updates.
Developer / integration partner — developers, technical partners, and IT teams need APIs, Docker, pack schemas, examples, and client snippets. Surface: setup guides, endpoints, package installs, integration templates, and validation commands.

Older phrases map to the canonical names as follows:

Older phrase	Canonical phrase
Social platforms, enterprise moderation, job marketplaces	Platform safety
NGOs, government bodies, consulates, regulators	NGO & regulator
Mobile/private checker, worker-facing app	Individual worker / mobile
Researchers, evaluators, Kaggle judges, reproducibility	Researcher
Knowledge exchange, anonymized knowledge sharing, signal exchange	Anonymized knowledge sharing
Developers, IT teams, custom integrations	Developer / integration partner

OFWs are a demo persona. Migrant workers globally is the category.

Public component names¶

Duecare Platform
│
├─ Gemma 4 Model Layer                 [Live]
├─ Safety Guidance Layer               [Live]
├─ Knowledge Packs                     [Live]
├─ Quality Testing Framework           [Partial]
├─ Local Anonymization Module          [Prototype]
├─ Information Submission Module       [Prototype]
├─ Central Knowledge Server            [Prototype]
├─ Public Information Research Monitor [Roadmap]
├─ Knowledge Formatter                 [Prototype]
├─ Stakeholder Engagement Module       [Roadmap]
├─ Newsletter and Alert Module         [Roadmap]
├─ Fine-Tuning Module                  [Prototype]
└─ Channel and Deployment Package      [Prototype]

Older internal names map to these public names: Runtime means Gemma 4 Model Layer; Harness means Safety Guidance Layer; Eval means Quality Testing Framework; Exchange means Central Knowledge Server plus Information Submission; Sentinel means Public Information Research Monitor; Trainer means Fine-Tuning Module; Channels means Channel and Deployment Package.

The live core for the Kaggle submission is Gemma 4 Model Layer + Safety Guidance Layer + Knowledge Packs + Quality Testing Framework (partial). Fine-Tuning is the appendix-notebook prototype. The mobile app is a sibling-repo proof for Individual worker / mobile. The central server, research monitor, stakeholder engagement, newsletter, and channel deployment modules are prototype / roadmap targets — their absence does not block the live demo.

1. Gemma 4 Model Layer¶

Model layer, not the whole product.

Responsibilities: generate plain-language explanations · summarize case facts · classify risk patterns · interpret screenshots and documents when multimodal is enabled · support local / private deployment · provide different model sizes for different surfaces.

Model tiers:

Tier	Best use
Gemma 4 E2B	mobile / on-device / low-resource worker checker
Gemma 4 E4B	NGO / government dashboard baseline
Gemma 4 26B / 31B	research, server-side eval, deeper reasoning
Fine-tuned Gemma	domain-specialized migrant-worker safety model
Multimodal Gemma	document screenshots, job ads, contracts, chat shots

Separation of concern: Gemma is the language and reasoning engine inside a deterministic safety system. The harness owns policy, rules, and evidence. Gemma does not invent laws, contacts, or policy.

2. Safety Guidance Layer¶

The core product. The current Kaggle submission is the first working implementation of this component.

Responsibilities: detect risk · retrieve context · structure evidence · ground model responses · prevent unsafe or ungrounded output · produce audit traces · provide deterministic, inspectable reasoning support around Gemma.

Submodules:

Submodule	Purpose
Anonymizer	Strip / generalize PII before storage or sharing
GREP / rule engine	Fast deterministic risk-signal detection (100+ rules)
RAG database	Laws, regulations, NGO guidance, pattern briefs (50+ docs + citation graph)
Tool layer	Fee calculators, jurisdiction lookup, corridor checks, and backing-table rows
Contacts directory	Official help channels, NGOs, regulators, consulates
Prompt / context builder	Converts signals + RAG into model context
Response policy	Bounds model behavior; no unsafe advice
Trace / audit layer	Shows why the system flagged something
Case summary layer	Produces reviewable intake notes

For every user group, the harness is the "why should I trust this?" layer. It turns Gemma from a chatbot into an auditable decision-support system.

3. Central Knowledge Server and Information Submission¶

This is what makes Duecare more than a static app. Hub scaffolded at duecare-ai.com — code in this repo at apps/duecare-ai.com/, deployed to Render via the repo-root render.yaml. The hub exposes POST /api/hub/signals (anonymized pattern-signal intake with a raw-PII rejector), GET /api/hub/knowledge-packs (knowledge-pack metadata listing), GET /api/hub/trends (aggregate counters), and GET /api/hub/status. The vetted-pack format + partner-discoverable index are roadmap; the shape of the proposal-and-distribution surface is live so partners can review the contract before participating.

Shared object types:

Shared object	Example
`RagDocumentProposal`	New regulation, labor advisory, court case, NGO report
`GrepRuleProposal`	New coded recruitment scam phrase
`RubricDimensionProposal`	New scoring criterion (e.g. contract substitution)
`ContactUpdateProposal`	Updated regulator hotline / complaint form
`AnonymizedCaseSignal`	"3 cases this month mention passport deposit + TikTok recruiter"
`EvaluationResult`	Model X failed 17 / 50 domestic-worker fee-manipulation prompts
`CorridorPatternUpdate`	New PH → Gulf recruitment pattern
`PolicyChangeNotice`	New law / regulation affects scoring

Privacy default: The Central Knowledge Server does not collect raw case data.

raw case
  ↓ local anonymizer
redacted signal / aggregate pattern / hashed provenance
  ↓ explicit partner opt-in
Central Knowledge Server
  ↓ human + automated validation
vetted knowledge pack
  ↓ pulled by clients
local harness update

Sharing-protocol invariants: no raw PII · no raw worker chats by default · all submissions have provenance · every update has contributor type + review status · every published pack is versioned and integrity-checked · clients can pin versions · sensitive cases stay local unless explicitly consented and anonymized.

Local cases create anonymized signals; public / verified sources create shared knowledge. Continuously updating becomes dangerous if it becomes centralized surveillance.

4. Quality Testing Framework¶

The research and quality-control layer.

Responsibilities: define dimensions · score model responses · compare models · validate new rules / RAG docs before release · prevent regressions · produce reproducible benchmark artifacts · support academic researchers and Kaggle judges.

Submodules: Universal rubric · domain-specific rubrics (trafficking, labor rights, tax evasion, financial crime, etc.) · evaluation questions (LLM-judge templates) · regression tests (prevent old failures from returning) · model comparison (Gemma vs other models, stock vs harnessed) · ablation tests (without RAG, without GREP, without fine-tune) · benchmark notebooks · governance checks (validate every new domain-pack release).

Role in the platform: Eval gates changes from the continuous update system. The agent proposes; Eval decides whether the proposal is safe enough to ship.

5. Duecare Trainer — model adaptation / retraining pipeline¶

Prototype (active A-00 pathway). The preconfigured pipeline at kaggle/A-00-omni-experiment-workbench/ is the active reference implementation for synthetic rows, LoRA smoke training, checkpoint/resume, adapter save/load, and report exports; full multi-tenant Trainer service is post-hackathon.

A privacy-preserving model-adaptation pipeline that converts validated, anonymized case knowledge into task-specific Gemma 4 adapters for moderation, case analysis, research evaluation, and worker assistance.

Adapter targets: duecare-gemma-evaluator (LLM-judge, prototype) · duecare-gemma-worker-advisor · duecare-gemma-case-intake · duecare-gemma-platform-moderation · duecare-gemma-multimodal-docs · corridor-specific adapters (-corridor-ph-hk, -corridor-id-sa, etc.) · duecare-gemma-research-assistant (last six all roadmap).

Hard rule — PII gate:

raw data → Anonymizer → curator validation → dataset builder
        → training (Unsloth LoRA) → Eval framework → release

No raw PII may enter the training set. Anonymizer's audit log stores sha256(original), never plaintext.

Why this matters across user groups: NGOs adapt models to their intake workflow; governments adapt to local regulations; platforms adapt to moderation policy; worker apps use smaller tuned models offline; researchers reproduce stock-vs-harnessed-vs-tuned behavior with provenance.

Composes with the Harness, does not replace it. A tuned model gives better baseline behavior; the Harness gives grounded, auditable, policy-controlled behavior. A judge or auditor inspects every layer regardless of which model loaded. Full architecture: architecture/duecare_trainer.md.

6. Public Information Research Monitor — continuous-update agent + server¶

Hub scaffolded; autonomous crawler is roadmap. The proposal-intake endpoint lives at POST /api/hub/opencrawl/updates on the public hub (duecare-ai.com, code at apps/duecare-ai.com/) for public-source crawler-style proposals. The autonomous crawler that fills that endpoint with continuously-discovered candidates is post-hackathon. When built, the crawler tracks laws, regulations, public reports, NGO guidance, platform abuse patterns, and partner feedback, then proposes vetted updates to Duecare's rules, RAG corpus, rubrics, contacts, and evaluation sets.

Agent functions: law / regulation tracker · NGO / stakeholder tracker · contact verifier (flags stale phone / email / form URLs) · RAG proposer · GREP proposer · rubric proposer · research-pack builder · release manager (builds vetted versioned packs after approval).

Human-in-the-loop is mandatory. The research monitor never auto-updates production safety behavior:

crawler / agent finds update
        ↓
proposal object
        ↓
automated validation
        ↓
human curator review
        ↓
Quality Testing Framework regression suite
        ↓
vetted pack release
        ↓
clients pull update

Tiered access:

Actor	Submit	Approve	Publish
Individual migrant worker	optional anonymized feedback	no	no
NGO partner	yes	maybe (own jurisdiction)	no / global limited
Government / regulator	yes	maybe (official contacts/laws)	no / global limited
Researcher	yes	benchmark / RAG proposals	no / global limited
Maintainer	yes	yes	yes

7. Channel and Deployment Package¶

Roadmap. Design target. Today's bridge: the chat package's FastAPI surface is the integration surface Channels would wrap with platform-specific 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.

Mobile is worker-owned. Channel deployments are institution-owned. Both use the same safety engine (Gemma 4 Model Layer + Safety Guidance Layer + Quality Testing Framework); 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 (in order of partner demand): web chat widget · WhatsApp Business · Facebook Messenger · Telegram · Viber · SMS (Twilio / Africa's Talking) · government website widget · NGO intake page (IFRC / IOM / IJM / Polaris / etc.).

Tenant knowledge pack per deployment combines global Duecare knowledge (ILO, Palermo, common indicators) + jurisdiction knowledge (local laws, complaint procedures) + institution knowledge (FAQ, intake scripts, escalation rules, human handoff contacts) + live updates from the research monitor via vetted Exchange packs.

Complaint-mechanism boundary is the same as the chat package and Mobile: explain options, identify likely contacts, prepare a draft, hand off to humans, let the user submit. Auto-submit / auto-email / auto-call / raw-message-to-server / hallucinated contact info / pretending-to-give-legal-advice are all banned by default. Canonical safety phrase used in every channel's session-start banner:

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

Submission disclosure. The Kaggle submission does NOT ship Channels. Auto-deploying on Messenger / WhatsApp without partner approval would create the impression that the bot is an official NGO channel when it isn't. Live deployment requires a named NGO tenant + business-account approval + Eval-gated tenant pack. Full architecture: architecture/duecare_channels.md.

8. Duecare Mobile — worker-facing private checker¶

The worker-facing product. Global migrant-worker focused, not OFW-specific.

A first implementation already exists in the sibling repo duecare-journey-android v0.9.0 with MediaPipe LiteRT on-device Gemma 4 E2B, encrypted SQLCipher journal, ILO indicators, and corridor packs. The roadmap items below are improvements to that existing app.

Modules:

Module	Purpose
Private checker	Paste / screenshot suspicious job offer or message
On-device Gemma	Local inference for privacy
Offline packs	Laws, contacts, warning signs by country / corridor
Multimodal scanner	Read job ads, contracts, screenshots
Risk explanation	Plain-language red flags
Trusted help	Official contacts and NGOs
Draft report	User-reviewed complaint / referral draft
Safety mode	"Do not confront recruiter / employer" guidance
Localization	Tagalog, Bahasa Indonesia, Nepali, Arabic, Hindi, Bengali, etc.

Critical boundaries:

Do not auto-report.
Do not tell workers to confront abusers.
Do not expose raw messages.
Do not store sensitive content by default.
Do not hallucinate contact info.
Do not pretend to provide legal advice.

The mobile app's framing line:

"This is a private risk check and resource guide. You decide what to do next."

How the six canonical lanes map to the eight components¶

Component	Platform safety	NGO & regulator	Individual worker / mobile	Researcher	Anonymized knowledge sharing	Developer / integration partner
Runtime	classify / explain posts	summarize cases	private explanations	compare models	sanitize candidate facts	embed model service
Harness	moderation risk trace	intake triage	warning signs	inspectable pipeline	reviewable knowledge-object flow	reusable API layer
Exchange	share anonymized abuse patterns	submit field signals	receive updated packs	publish domain packs	primary lane	integrate safe submissions
Eval	policy QA / model QA	quality control	hidden safety guard	primary tool	PII and provenance gate	regression gate
Trainer	moderation adapter	agency-specific adapter	smaller local model	tuned model studies	approved sanitized examples only	BYO adapter path
Research monitor	track new scam patterns	update laws / contacts	keeps app current	benchmark updates	candidate source context	public-source proposal API
Channels	platform notices	primary deployment	Messenger / WhatsApp help	study artifact	reviewer submission surface	integration target
Mobile	not primary	referral companion	primary worker-owned app	deployment study	opt-in aggregate signals	app integration pattern

Every component serves multiple audiences.

Separation-of-concerns rules¶

Models do not own truth. Gemma generates and reasons; it does not invent laws, contacts, or policy. Trusted data comes from RAG packs, contact packs, rule packs, and reviewed proposals.
The update agent proposes; humans approve. The research monitor does not silently mutate production safety behavior.
Raw cases stay local. Exchange shares anonymized signals, not raw worker narratives.
Contacts are deterministic. Contacts come from verified data, not model output.
Evaluation gates every update. New rules / docs / rubrics run through regression tests before release.
Mobile app consumes vetted packs. The app pulls versioned, offline-compatible safety packs.

High-level diagram¶

                               Public + partner sources
                 laws, regulations, NGO docs, contacts, case learnings
                                             │
                                             ▼
                         Public Information Research Monitor
                  crawl → propose → validate → human review → release
                                             │
                                             ▼
                              Vetted Duecare Knowledge Packs
               RAG docs | GREP rules | contacts | rubrics | eval prompts
                                             │
        ┌────────────────────────────────────┼────────────────────────────────────┐
        ▼                                    ▼                                    ▼
 Fine-Tuning Module           Safety Guidance Layer        Channel and Deployment Package
 LoRA + GGUF + LiteRT         GREP + RAG + tools + tests    Messenger / WhatsApp / web
        │                                    │                                    │
        └────────────────────────────────────┼────────────────────────────────────┘
                                             ▼
                                   Gemma 4 Model Layer
                                             │
                      ┌──────────────────────┼──────────────────────┐
                      ▼                      ▼                      ▼
              Platform safety          NGO & regulator       Individual worker / mobile
                                             │
                                             ▼
                                     Quality Testing Framework
                              reproducibility, regressions, benchmarks

What this means for the current Kaggle submission¶

The Kaggle submission is the first working implementation of the Safety Guidance Layer, running on the Gemma 4 Model Layer, with a partial Quality Testing Framework baked in and a Fine-Tuning Module prototype shipped through the DueCare Fine-tuning and Evaluation.

Real now¶

Gemma 4 Model Layer — Gemma 4 model selector across local and cloud/BYOK targets, output sanitizer.
Safety Guidance Layer — 100+ GREP rules, 50+ document RAG, function-calling tools, and contact packs.
Quality Testing Framework (partial) — multi-dimension rubric + adversarial suite + multiple grade modes.
Contacts directory — curated entries for regulators, NGOs, embassies, and hotlines.
Synthetic multimodal evidence — bundled CC0 images + structured-post JSONs.
Kaggle live proof path — active kernels: exploration workbench, live demo, and A-00 Fine-tuning and Evaluation proof material.

Prototype / appendix / near-term¶

Fine-Tuning Module — active kaggle/A-00-omni-experiment-workbench/ runs synthetic row generation, Unsloth LoRA smoke training, checkpoint/resume, adapter save/load, four-arm comparison, combined judging, and report exports. Adapter targets are scoped; multi-tenant fine-tuning service is post-hackathon.
NGO / government RAG packs — _contacts.json curator-block pattern is the bridge. Tenant-specific packs are post-hackathon.
Complaint draft workflows — /static/hotlines.html + /api/contacts produce mailto: drafts. Channel-specific UX (Messenger / WhatsApp) is roadmap.

Roadmap¶

Central Knowledge Server and Information Submission — privacy-preserving vetted-pack distribution. Today's bridge: curator-block PRs reviewed by humans.
Public Information Research Monitor — continuous public-source update server. Today's bridge: scripts/v141_validate_contacts.py is a read-only HEAD-pinger seed.
Channel and Deployment Package — Messenger / WhatsApp / web / SMS / embassy-portal adapters. Today's bridge: the chat package's FastAPI surface is the integration surface; no live channel deployment.
Individual worker / mobile app v1.0 — Duecare Journey v0.9.0 ships today; v1.0 consumes vetted packs, expands multimodal screenshot scanner, expands corridor coverage.

This is a strong "real now + scalable vision" story: functional core today, credible platform vision tomorrow. The Kaggle submission ships the live core (Gemma 4 Model Layer + Safety Guidance Layer + Quality Testing Framework + Contacts) and an appendix prototype (Fine-Tuning Module); the Central Knowledge Server, Public Information Research Monitor, and Channel and Deployment Package are documented design targets that partners can review without confusion about what's running today vs. what's planned.