Migrant worker safety AI research project

A Gemma 4 safety ecosystem for migrant-worker protection.

DueCare is a thin, model-agnostic legal-grounding layer for migrant-worker protection. Before a large language model answers a worker’s question, DueCare fires deterministic ILO-indicator rules, retrieves the governing law, and runs verified tool calls, then adds that grounding to the prompt and nothing else. The result is measurably safer answers about recruitment fees, passport retention, and contract substitution, on the open models NGOs, platforms, and workers already run.

Who it's for: Platforms/ NGOs & regulators/ Individual worker / mobile/ Researchers/ Anonymized knowledge sharing/ Developers
▸ Watch the demo → Get started See use cases Harness ecosystem Download Android app
Walkthrough Full walkthrough
A guided tour of DueCare, end to end.
/duecare/demo Play

Six lanes, one runtime

DueCare ships six audience-facing lanes — content moderation, case analysis, worker support, research, anonymized knowledge sharing, and custom API implementations — all sharing the same Gemma 4-centered safety stack.

01Platform safety

Screen exploitative user-generated content, at scale.

Trust & safety teams, recruitment marketplaces, and high-volume UGC review queues.

Detect illegal fees, contract substitution, passport retention, deceptive ads, and other exploitative UGC patterns. The system flags and explains; your existing review pipeline acts.

→ read more
02NGO & regulator

Speed up case analysis and triage.

Caseworkers, labor regulators, embassies.

Run on the caseworker's own device. Surface the relevant laws and advisories, draft replies and complaint forms, and flag patterns across cases, without case data leaving local machines.

→ read more
03Individual worker / mobile

Plain-language answers about a corridor.

Migrant workers, community channels, partner-hosted endpoints.

Answer questions about fees, contracts, passports, and rights in the worker's own language. The system never files complaints or instructs action; it points to verified resources and trusted caseworkers.

→ read more
04Researcher

Cite-able research on corridors, trends, and policy.

Academic researchers, policy analysts, journalists.

Study corridor risk, exploitation trends, and policy impact using version-pinned packs and an anonymized signal stream. Run the evaluation and research harnesses on Gemma 4; cite a hash; re-run months later.

→ read more
05 Anonymized knowledge sharing

Share verified facts, not raw case files.

Reviewed, redacted facts become reusable knowledge objects after local sanitization and human approval. Partners can improve corridor packs without exposing workers, private documents, names, contact details, or case narratives.

  • Reviewed evidence graphs
  • Local PII anonymization
  • Consent-aware metadata
  • Vetted knowledge objects
→ sharing flow
06 Developer / integration partner

Drop the runtime into your own product or channel.

A containerized runtime with a small, well-defined API. Wire it into a Messenger or WhatsApp bot, a moderator console, an internal case-management system, or a sibling app. You own the channel; we provide the harness ecosystem, the rule packs, and the model layer.

  • FB Messenger / WhatsApp adapter
  • Custom moderator dashboards
  • Case-management plug-ins
  • On-prem deployments
→ integration paths
One worker, one question

The gap is one grounded answer wide.

The same model, asked the same question, with and without the grounding layer. The bare answer sounds reasonable. The grounded answer flags the indicator, cites the law, and points to a verified human.

Composite worker · illustrative
“A recruiter in Kathmandu says the Gulf job is guaranteed, but I have to pay NPR 150,000 up front and hand over my passport when I land. Is that normal?”
Ramesh · Nepal → Gulf domestic-work corridor

What the grounding layer adds

  1. Flags the fee as an indicator
    An ILO-indicator rule fires on “pay up front”: charging recruitment fees to the worker is a forced-labour indicator under ILO C181. The bare model had called it “sometimes standard.”
  2. Cites the corridor rule, not a guess
    Retrieval pulls the corridor pack: in this corridor the employer bears recruitment costs, not the worker. The answer quotes that rule with its source.
  3. Names the passport risk and a verified route
    Surrendering a passport on arrival is flagged as document retention. A tool call resolves a current regulator or NGO contact, so the referral is verified, never invented.
Measured, not asserted

A thin grounding layer, and every model got safer.

DueCare adds fired ILO-indicator rules, retrieved law, and deterministic tool calls to the prompt, and nothing else. An independent three-model judge panel then scores each answer, bare versus grounded, 0–100 across five safety dimensions. The paired difference is the harness lift.

Mean harness lift
+40.7 / 100
gemma4:31b · 48.4 → 89.1 component judge
Prompts improved
99.8%
15 of 7,953 scored worse
Paired prompts
7,953
one model, answered twice, judged twice
Judge panel
3-model
independent; never grades its own family

The lift holds across every model tested — 8 models, positive on every one (paired Cohen’s d = 1.73). Because the grounding is added at the prompt, it works the same on a model you already run.

How it works

01

Fire the rules

Deterministic ILO-indicator rules scan the message before the model speaks — recruitment fees, passport retention, contract substitution, and more.

02

Ground the answer

Retrieve the governing corridor law and vetted facts; deterministic tool calls resolve fee caps and verified hotlines. All of it is added to the prompt.

03

Measure the lift

An independent three-model judge panel scores each answer bare versus grounded, 0–100 across five safety dimensions. The paired delta is the lift.

What this does — and does not — prove

What the evidence shows
  • Grounding improves response quality, measured by an independent LLM-judge panel on a paired design that cancels judge bias.
  • Scores are computed on held-out prompts the training never sees, so the lift is not memorization.
  • The benchmark is fact-checked against its own data — 20 of 20 headline claims reconcile — and added citations verify at a 0% hallucination rate.
What it does not claim
  • Not a measure of real-world trafficking detection — it is benchmark evidence about answer quality, not ground truth about any person.
  • The judges are language models, not anti-trafficking professionals; a blinded human-expert validation is the honest precondition for a peer-reviewed claim.
  • Raw worker chats, IDs, and documents stay on-device; only anonymized, pre-approved envelopes ever leave.
See the full results → Download the data Live leaderboard Fine-tuning methodology
/ use cases

See the full set of use cases & deployments.

Reference scenarios for Platform safety, NGO & regulator, Individual worker, Researcher, Anonymized knowledge sharing, and Developer / integration partner lanes, with the deployment architecture for each one in a single place.

View use cases →
01 · Platform safety

Screen exploitative user-generated content, at scale.

Platform safety teams and recruitment marketplaces face a constant stream of posts, ads, listings, and messages, some of which target migrant workers with illegal fees, deceptive offers, or unsafe employers. The DueCare runtime classifies and explains UGC against curated corridor packs, returning a risk score and a suggested action your existing enforcement pipeline can act on.

What it detects

Illegal recruitment fees, contract substitution, passport retention requests, deceptive job ads, sham agencies, and other exploitative UGC patterns.

What it returns

A per-item risk score with a suggested action: hold, redline, escalate to human review, or notify the relevant authority.

Where it runs

Inside the platform's own VPC, on a self-hosted GPU. No user content leaves the platform environment.

What it does not do

It does not auto-enforce. The platform's existing trust & safety pipeline decides what action to take.

→ See the platform deployment architecture
02 · For NGOs & regulators

Speed up case analysis and triage.

Caseworkers and labor regulators handle complex situations under time pressure. Running on the caseworker's own device, DueCare surfaces relevant laws and advisories, drafts replies and complaint forms, and flags patterns across recent cases.

How it helps analysis

Pulls the right corridor pack sections automatically; suggests follow-up questions; pre-fills complaint or referral forms with citations.

How it helps triage

Flags patterns across recent local cases (e.g. a recurring fee scheme) without those cases ever leaving the device.

Where it runs

On the caseworker's laptop or workstation. No case data leaves local machines.

What it does not do

It does not file complaints, contact employers, or take any external action on the worker's behalf.

→ See the NGO deployment architecture
03 · Individual worker / mobile

Local mobile help. For the worker’s device and language.

Migrant workers need answers now, on the device or channel they already use, in the language they actually speak. The worker/mobile lane uses cached corridor packs, local or partner-controlled inference, and a no-raw-upload boundary to the public hub. It points the worker to verified resources rather than acting on their behalf.

How workers reach it

Through a partner-operated channel: an NGO chatbot, a government information line, or a sibling worker app.

What it answers

Questions about typical fees, legal contract terms, passport retention rules, and where to seek verified help in this corridor.

Where it runs

On a worker-controlled device when an on-device build is available, or through a partner-hosted endpoint where the partner controls the channel and audit trail.

What it does not do

It will not file complaints, instruct action, or replace a verified caseworker. Drafted answers always point to verified humans.

→ See the chatbot deployment architecture
04 · Researcher

Cite-able research on corridors, trends, and policy.

Academic researchers, policy analysts, and journalists study migration corridors, exploitation trends, recruitment-market dynamics, and policy impact. DueCare gives them version-pinned packs, an anonymized signal stream, and a reproducible harness on Gemma 4 so claims about how a corridor is moving, or how a policy change landed, can be cited and re-run months later.

What it supports

Corridor risk analysis, longitudinal trend tracking (fees, contract substitution, recruiter networks), policy-impact studies, and reproducible model evaluations.

What is reproducible

Pack version (cryptographic hash), eval prompts, rubric, model artifact, runtime configuration. plus the anonymized signal aggregates queried at a point in time.

Where it runs

Locally on Gemma 4 or in a cloud notebook. Researchers control their own environment; the harness reads only public packs and aggregate signals.

What it does not do

It exposes no individual workers, employers, or cases. Aggregates are k-anonymized; raw case data never leaves local deployments.

→ See the research deployment architecture
05 · Anonymized knowledge sharing

Share verified facts, not files.

The sharing lane turns reviewed, redacted evidence into reusable knowledge objects. Case files stay inside the local workbench; only sanitized, consent-aware facts with provenance and reviewer approval can reach the public hub.

What gets selected

Confirmed facts from reviewed evidence graphs: corridor, indicator, rule, source row, outcome, and pack version.

What gets removed

Names, phone numbers, IDs, document images, private narratives, employer-identifying details, and other sensitive PII.

Where it runs

Sanitization runs in the local or trusted deployment before any submission. The hub runs a second defensive PII scan before storage.

What improves

Future packs, GREP rules, worker guidance, platform moderation, and research queries improve from verified patterns without centralizing raw cases.

→ See the anonymized knowledge sharing flow
06 · Developer / integration partner

Drop the runtime into your own product or channel.

The runtime is a containerized service with a small, well-defined API. Wire it into a Messenger or WhatsApp bot, a moderator console, an internal case-management system, or a sibling app. You own the channel; we provide the harness ecosystem, the rule packs, and the model layer.

Channel adapters

FB Messenger, WhatsApp Business, Telegram, custom web chat, voice IVR. Adapter handles channel-specific transport; the runtime stays the same.

Console integrations

Custom moderator dashboards, in-house case-management plug-ins, audit log exporters.

Deployment shape

On-prem, in your VPC, or partner-hosted. Containerized; runs on standard GPU hardware.

What we provide

The runtime image, the harness configuration, the rule pack catalog, and an API client. You provide the channel.

→ See all deployment paths