No-wait recording deck

Five system surfaces in under three minutes.

This page is the recording spine: it lets the video show the real DueCare architecture without waiting for model inference. The live app can still be opened in adjacent tabs, but the primary take uses cached, already-approved response examples and short screen jumps.

0:00-0:25

Human stakes

Composite worker story and privacy promise.

0:25-0:55

Private check

On-device worker/NGO response with GREP + RAG + tools.

0:55-1:25

Multimodal analysis

Document photo and case-packet extraction.

1:25-2:05

Public hub

Knowledge packs, submissions, and admin logs.

2:05-2:55

Training loop

Harness traces become SFT + DPO data, then gated release.

Shot list

The five fast cuts

Scene 01 · Impact

Composite worker risk check

Show the Individual worker claim: private, local, and actionable.

cached response · worker checkerno inference wait
GREP firedRAG citedtool draft
Risk: high. The message combines recruitment-fee pressure, identity-document control, and debt leverage.
Safe answer: do not hand over a passport or pay new fees without a written receipt and verified legal basis. Contact a trusted NGO, consulate, or labor office before replying.
Why it matters: the model refuses to help optimize the recruiter script and instead routes the worker to safe, local resources.
Scene 02 · Multimodal

Document and case-packet analysis

Show Gemma 4-style document understanding as a real workflow: photo/text in, extracted facts out.

cached response · document inspectormultimodal fallback safe
passport clausefee amounttimeline
Extracted facts: placement fee demand; passport handover clause; wage-deduction schedule; recruiter agency reference.
Generated packet: timeline, legal-grounding notes, questions for the caseworker, and draft complaint outline.
Video move: show the extracted-fact table, not a spinning model call.
Scene 03 · Public hub

Knowledge packs and safe coordination

Show that one NGO’s anonymized knowledge can help another NGO or regulator without moving case data.

cached response · hub syncpack registry
4 example packssigned versionspublic-source only
Pack pulled: phl-kwt-domestic@1.7.2
Filters: ContextPack · PHL-KWT · vetted
Update path: partner proposal → PII gate → automation triage → curator review → vetted pack.
Scene 04 · Admin

Troubleshoot the website during the recording

Use the token-gated admin page to confirm submissions, updates, local-KB counts, and JSONL storage health.

cached response · admin logstoken gated
signals.jsonlupdates.jsonlredacted output
Admin view shows: last 50 accepted signals, update proposals, automation verdicts, file byte counts, counters, and local-KB case count.
Safety rule: free-form payloads are suppressed and detector-class PII is redacted before display.
Scene 05 · Training loop

Harness traces become model improvement data

Close on technical depth: SFT + DPO, with evaluation gates before release.

cached response · A07 training spineno RL claim
SFTDPOrelease gate
Chosen answer: Persona + GREP + RAG + Tools response.
Rejected answer: raw Gemma answer when it misses law, safety, or citation grounding.
Dataset: prompt, chosen, rejected, provenance, rubric grade, and gate metrics. This is preference optimization, not a fake live RL loop.
Priority examples

Hand-selected scenarios to pre-run before the final cut

The page pulls a synthetic no-PII catalog from /api/demo/priority-examples. For the final video, keep these IDs and replace the fixture outputs with warmed runtime outputs from Gemma 4 + the DueCare harness so the take stays fast and real.

Loading

Fetching priority examples…

If JavaScript is disabled, open /api/demo/priority-examples directly.

Methodology

How to capture without waiting on inference

Primary take

Record this deck plus the live hub pages. The cached response blocks are the spoken proof points; the live routes are the visual proof that the product exists.

  1. Open this page, /demo, /knowledge-packs, /admin, and /evaluation.
  2. Pre-load all tabs before recording.
  3. Use hard cuts instead of waiting for network or inference.
  4. Only show live inference as a short optional B-roll clip if it is already warm.

Backup notebook

If Kaggle is the required proof surface, mirror these same five scenes into a cached-response notebook: markdown explanation, JSON constants, and rendered tables. It should not load a model; it should cite the artifact path or API route that produced each cached response.

  1. Precompute responses once from the wheel/runtime.
  2. Commit only synthetic/composite examples.
  3. Run the notebook before screen capture so outputs are already visible.
  4. Use the live app for proof, the notebook for narration pacing.