Skip to content

For Hackathon Judges - Verify DueCare in 5 minutes

Canonical reviewer entry doc: FOR_PEER_REVIEW.md. This page is the hackathon-specific quick path. The peer-review doc is the broader verification roster aimed at any reviewer (academic, journalist, policy analyst, NGO technical lead, or hackathon judge). Both link to the same kernels, packs, and live surfaces; this doc focuses on the 5-minute hackathon-rubric verification flow.

The DueCare submission for the Gemma 4 Good Hackathon (2026-04-02 → 2026-05-18). This doc exists so you don't have to spelunk the repo to verify any claim — every claim below points to a specific file, kernel, or live URL.

In a hurry? Run kaggle/02-live-demo/kernel.py, open /start, and click Project slides. The deck works without a model; live Gemma 4 is used for chat and optional Bulk File Review edge creation.

What it is, in one paragraph

DueCare is a Gemma 4 safety ecosystem for migrant-worker protection. This submission has two active Kaggle kernels: the DueCare App workbench and the DueCare Live Demo. Together they prove the same end-to-end system: Gemma 4 runtime, safety harness, Bulk File Review case analysis, graph extraction, anonymized knowledge sharing, and contact routing. The active A-00 path preserves the stock/fine-tuned/harness evaluation matrix. The public hub at duecare-ai.com is the companion website for the judges-facing story, kernel links, knowledge-pack registry, and anonymized signal intake. Code for the hub is in apps/duecare-ai.com/ and deploys to Render via the repo-root render.yaml.

The harness wraps Gemma 4 with retrieval (BM25 + optional hybrid dense-retrieval), structural document chunking, 451 hand-curated trafficking-pattern rules, an 865-document RAG corpus + citation graph (counts regenerable via python scripts/verify_knowledge_surfaces.py), and a multi-mode grading stack — turning stock Gemma into a domain-specialised LLM safety judge for migrant-worker trafficking scenarios. The six canonical lanes are (1) Platform safety, (2) NGO & regulator, (3) Individual worker / mobile, (4) Researcher, (5) Anonymized knowledge sharing, and (6) Developer / integration partner. Those six lanes support three outcomes: prevent exploitation before it spreads, assist victims and at-risk workers, and help stakeholders understand what is happening and why. Filipino overseas workers are one demo persona, not the product category. Built specifically for partners who cannot send sensitive case data to frontier APIs, with validated local/demo surfaces now and a sibling DueCare Journey Android APK proving the LiteRT/offline edge path.

Notebook status note: the current judge-facing Kaggle path is the active kernel set in kaggle/_INDEX.md: exploration workbench, live demo, and DueCare Fine-tuning and Evaluation (the A-00 quantitative proof kernel, taylorsamarel/duecare-fine-tuning-and-evaluation). Retired notebook-era appendix surfaces are historical.

Where to verify each claim

Claim Verify at
Multi-dimension universal rubric packages/duecare-llm-chat/src/duecare/chat/harness/_rubric_universal.json
451 GREP rules across active categories (crypto / scam-compound / gig-economy / BNPL / Ukrainian + Afghan corridors / Pacific RSE-PALM / sub-Saharan / EU posted workers / climate-displaced / intra-community / domestic-to-sex-work transition / cross-platform signals) packages/duecare-llm-chat/src/duecare/chat/harness/__init__.py (GREP_RULES = [...])
Large example prompt library across audience buckets (model_capability / enterprise_moderation / ngo_intake / individual_query / research / image_prompts / data_intelligence / regulator_audit) packages/duecare-llm-chat/src/duecare/chat/harness/_examples.json
Bundled CC0 synthetic-evidence images + structured-post JSONs with watermarks + sidecar JSONs + cross-platform-signal links packages/duecare-llm-chat/src/duecare/chat/static/synthetic/ + static/synthetic/posts/
865-document curated RAG corpus (ILO C29/C95/C97/C143/C181/C188/C189/C190/P029, fair-recruitment guidance, POEA/BP2MI/Nepal/HK/SG/UAE statutes, WHO Global Code, EU 2024 ATD amendment, ASEAN ACTIP, CoE 197, CEDAW GR 38, UNCRC, UNODC/IOM/FATF typologies, Pacific Climate Mobility, Bali Process) packages/duecare-llm-chat/src/duecare/chat/harness/__init__.py (RAG_CORPUS = [...])
Citation graph for 1-hop graph expansion at retrieval time packages/duecare-llm-chat/src/duecare/chat/harness/_citations.json
Adversarial validation suite across multiple attack families including new structured-data-attack / image-injection / regulator-impersonation / multilingual-jailbreak scripts/adversarial_validate.py + reports/adversarial_*.md (run output)
evaluator_call hook for separating LLM-judge model from chat model (v0.13.0; abliterated / frontier / larger-Gemma patterns supported, Kaggle-default in-process self-grade due to VRAM) packages/duecare-llm-chat/src/duecare/chat/app.py:_evaluator_model_call
Hybrid retrieval — BM25 + optional dense + RRF fusion, with reranker hook app.py:_hybrid_fuse_with_dense + kernel_helpers/embedding.py
Path tracing — every retrieval-pipeline stage logged app.py:_path_trace_record + chat UI's "RETRIEVAL PATH TRACE" card
A/B Compare tab — same prompt, two harness configurations, side-by-side grades kaggle/01-duecare-exploration-workbench/kernel.py live; UI button next to "About"
License + attribution — every bundled asset, every model, every third-party reference LICENSES.md
Competition rule compliance — winner licence, external-data reasonableness, reproducibility deliverables, open-source and warranty provisions, mapped section by section COMPETITION_COMPLIANCE.md
Responsible use policy — what DueCare is and is not, prohibited uses, why the adversarial corpora are released tiered, privacy boundaries, reporting routes RESPONSIBLE_USE.md
Public hub — knowledge-pack registry + anonymized signal intake + public-source proposal intake duecare-ai.com (code at apps/duecare-ai.com/, deployed via repo-root render.yaml)
Hub API surface — GET /api/hub/knowledge-packs, POST /api/hub/signals, POST /api/hub/opencrawl/updates, GET /api/hub/trends, GET /api/hub/status, GET /api/health duecare-ai.com/docs (FastAPI auto-generated OpenAPI)
Recording-grade deck and demo routes kaggle/02-live-demo/kernel.py prints a *.trycloudflare.com tunnel; open /start, /slides, and /wb-static/process.html
Video script and recording path docs/video_script.md plus /slides/setup or scripts/prebake_slide_cached_io.py for the cached worker-question row
Fine-tuned adapter evidence exported by the A-00 pipeline when training is enabled; see /kaggle/working report and manifest
LiteRT / on-device path sibling DueCare Journey Android APK and bundled harness manifest

Hackathon track qualification

Competition overview alignment checked on 2026-05-18 against public Gemma 4 Good summaries: the project should be filed under Impact Track -> Safety & Trust, with Special Technology evidence for Unsloth and LiteRT. The same overview emphasizes local frontier intelligence, native function calling, multimodal understanding, a working demo, public code, technical analysis, and a 3-minute video; the repository maps each requirement to a concrete surface below.

Track Why this submission qualifies Anchored to
Impact Track → Safety & Trust Core concept is LLM safety for the most vulnerable migrant-worker populations. The harness directly targets known trafficking-detection failure modes. The 50-prompt adversarial suite + the harness-on-vs-harness-off A/B comparison.
Special Tech → Unsloth Fine-tune is via Unsloth + LoRA on Gemma 4. A-00 preserves the training config, checkpoints, adapter path, and comparison report. kaggle/A-00-omni-experiment-workbench/kernel.py + exported A-00 report bundle
Special Tech → LiteRT The sibling Android app bundles harness metadata and uses the LiteRT path for the worker-facing offline surface. github.com/TaylorAmarelTech/duecare-journey-android (+ /releases for the APK artifact built by GitHub Actions)
Main Track Pursued in parallel; depends on overall execution + video. All of the above bundled.

Five-minute walkthrough

Run kaggle/02-live-demo/kernel.py or open the current Cloudflare URL printed by that kernel. The route sequence below is stable:

  1. Open /start. Confirm the two-tile landing: Project slides and Workbench.
  2. Open /slides. Walk the 23-slide deck. The demo slides use cached replay controls, so a recording does not wait on GPU output.
  3. Open /slides/setup if you want to refresh the cached worker chat row, or use scripts/prebake_slide_cached_io.py to preload it.
  4. Open /wb-static/process.html. Drop case_files_streamlined_demo.zip, then watch upload, processing, review, graph creation, and optional local Gemma 4 edge generation.
  5. Ask a graph question after confirming extracted intelligence. The answer should cite row IDs, evidence edges, and extracted entities from the local staging directory.
  6. Open the DueCare App kernel for the full chat harness and A/B comparison surfaces.
  7. Optional: open the A-00 report to verify the stock, harness, fine-tuned, and fine-tuned-plus-harness evaluation matrix.

If you can do all 7 steps, you have empirical verification of the writeup's live-demo, local-processing, and evaluation claims.

Design note: which model grades the LLM-based scoring?

The chat package's "LLM-Based" and "Combined" grading modes (G-Eval / MT-Bench / Auto-J style) ask a model one yes/no question per rubric dimension. In this Kaggle deployment, that is the SAME Gemma instance loaded for chat — an on-device self-grade. The deployment constraint that forces this choice is real and worth naming explicitly:

Kaggle T4 / dual-T4 sessions are VRAM-bounded. The loaded chat model (Gemma 4 E2B / E4B / 26B-A4B / 31B variants) saturates VRAM at load time. Holding a SECOND model in memory for grading (abliterated variant, frontier-via-API stub, larger Gemma) would require unloading the chat model mid-session — slow, breaks user state, and undermines the live-demo flow that judges click through. For the hackathon Kaggle deployment we therefore run in-process self-grade by necessity, not by design preference.

The chat package nevertheless treats the chat model and the LLM-judge model as independently-configurable hooks (the evaluator_call parameter on create_app, v0.14.2+). The hook is documented architecture for production deployments outside the Kaggle memory constraint where you have headroom for two model loads or a network call to a separate judge process. Three production patterns the hook supports — commentary only on this submission since none of them run in the Kaggle session:

  1. Frontier judge (GPT-4 / Claude 3.5 / Gemini 1.5 Pro) — gold- standard accuracy via API. The chat model stays on-device for the privacy-preserving Individual worker use case; the judge runs separately, the G-Eval / MT-Bench / Auto-J methodology. Network requirement means it doesn't fit the on-device-only Kaggle exercise.
  2. Purpose-built safety classifier (e.g. Google's ShieldGemma family) — scores harmful content as a classification task, so it returns a grade where a chat-tuned model would decline to engage with the text at all. This is the correct answer to the "the judge refuses to grade the adversarial suite" problem: a model built to classify harm, not a model with its refusal training removed. DueCare does not use, endorse, or distribute refusal-ablated Gemma weights.
  3. Larger Gemma (e.g. Gemma 4 31B-it grading while chat runs E2B) — better grading without giving up the chat-side speed. Same VRAM-headroom precondition.

The Kaggle-deployed kernel ships with evaluator_call=None (the default) so the judges-time experience matches what the writeup claims. The hook stays as documented architecture for the post-hackathon production-deployment story. To swap (in a non-Kaggle deployment with VRAM headroom), one line in kernel.py:

app = create_app(
    gemma_call=loaded.backend,             # E4B for chat
    evaluator_call=heavier_loaded.backend, # 31B-abliterated for grading
    # NOT advisable on Kaggle — would OOM the session.
    ...,
)

What this submission deliberately does NOT do

  • It does not bundle real-person evidence imagery. Every image shipped in the wheel is auto-generated CC0 synthetic with visible watermark + composite character names + reserved-for-fictional-use phone-number prefixes. See LICENSES.md for the synthetic- disclaimer.
  • It does not claim to detect every trafficking pattern in the world. 451 GREP rules + 865 RAG docs + 46 citation edges cover the major documented vectors well, but new patterns appear weekly — the curator-block JSON pattern lets stakeholders contribute updates without a code change.
  • It does not replace human caseworkers. The harness's design explicitly routes to named NGOs / regulators / embassies rather than offering itself as the answer.

Reading order if you have 30 minutes

  1. docs/writeup_draft.md — the 1,500-word submission writeup
  2. LICENSES.md — full attribution + license declarations
  3. docs/COMPETITION_COMPLIANCE.md — rule-by-rule compliance map (winner licence, external data, reproducibility, warranty)
  4. docs/USER_TODO.md — author's submission-day checklist
  5. reports/adversarial_<latest>.md — the empirical evidence
  6. The current Cloudflare URL printed by kaggle/02-live-demo/kernel.py

Reading order if you have 5 minutes

  1. This document
  2. The Kaggle live demo URL (technical depth)
  3. duecare-ai.com (platform infrastructure)
  4. The video