FILE Health-product integrity · prepared for HSA

Check it
before you
swallow it.

A public-safety instrument for Singapore: verify any health-product listing before you buy, and turn every check into enforcement signal. A ScamShield-class channel, built for HSA's regulatory domain.

Extension-firstOn-deviceBedrock · ap-southeast-1Sovereign by design
14
HSA alerts in the blacklist
14
banned substances tracked
4
e-commerce platforms covered at launch
4
languages screened — EN · ZH · MS · TA
0
shopper identifiers leave the device
REC. 01 The brief

Fakes moved to the feed. Enforcement hasn't caught up.

Counterfeit, adulterated, and illegally-sold health products circulate freely on Shopee, Lazada, TikTok Shop and Carousell. HSA's own advisories name the casualties: slimming pills laced with banned sibutramine that put a consumer in hospital; counterfeit probiotics that sickened a family; dietary supplements spiked with undeclared potent drugs.

Takedowns work, but they are reactive. A listing is pulled after harm is reported, then reappears under a new name the next week. Consumers have no way to check a product at the one moment that matters: before they tap buy.

The people most often conned, the elderly, the unwell, the hopeful, are the least equipped to spot a fake.
REC. 02 The insight

You can't verify what was never registered. So don't pretend to.

Health supplements aren't pre-approved by HSA, so there is no “genuine products” list to check against. The honest architecture is three classes, and the discipline is never dressing an estimate up as a guarantee.

Class I · Known-bad
Deterministic block
A match to an HSA alert or a banned ingredient is certain. The system says stop, and says exactly why. No model in the loop.
Class II · Registered
Confirmed listing
Medicines, CPM and devices carry real registrations. We confirm the listing exists: “registered,” never “genuine” or “endorsed.”
Class III · Unknown
Explainable risk score
For the long tail, a scored assessment with named reasons. Never an authenticity guarantee, and cold-start returns “insufficient data,” not a green light.
APPENDIX A Detection methods

Each layer defeats a specific move.

Fraud adapts, so the engine is built around the tactics, renaming, hidden ingredients, in-image claims, stolen photos, price bait, review farming, not around generic “AI.”

Tactic defeatedDetection method
Renaming after takedown
Semantic blacklist
Every HSA alert is embedded, so a re-skinned banned product matches on meaning even when the name and packaging change.
RAGpgvector
Claims hidden in images
Vision reading
A vision-language model reads product images, catching miracle and disease-cure claims that never appear in the listing text.
Bedrock ClaudeVLM
Undeclared adulterants
Ingredient NER
Multilingual extraction pulls ingredients from EN, ZH, MS and TA text and checks them against HSA's prohibited-substance list.
MultilingualNER
Brand-jacking with stolen photos
Image similarity
Perceptual hashing plus multimodal embeddings spot stolen brand photography and known-counterfeit packaging in circulation.
Cohere Embed v4pHash
Price bait
Price anomaly
Listings cluster to a reference price; implausible discounts become a measured outlier score instead of a gut feeling.
Robust z-scoreClustering
Review farming
Review red-flags
Velocity bursts, U-shaped ratings, thin profiles and LLM-written reviews, plus adverse-event mining that surfaces real harm.
Burst detectionSignals
SCHEMATIC How it works

Extraction on the edge, judgment in the region.

Detection runs on listing facts, not on the shopper. Everything that can happen on the user's own device does, and only a feature vector ever reaches the service.

On device
Extract & gate
Reads the listing in the user's session, strips identifiers, instant local block on known-bad.
Edge → region
Feature vector
Only listing facts travel, never the page, cart, or identity.
ap-southeast-1
Tiered engine
Deterministic block, registry match, scored signals, aggregated and explained.
Back to user
Explainable verdict
A classification, a risk index, and the specific reasons, in context and in page.
Phase 2
HSA triage
Officers adjudicate reports into the blacklist, protecting every user at once.
↻Every check and report feeds the seller graph and the officer queue. Confirmed cases propagate back to all users, the same flywheel that lets ScamShield block over 120,000 entities.
FIELD USE In context

Where the buying happens.

On desktop the classification appears in-page as you browse. In native apps, where most Singaporeans actually buy, one tap of Share sends the listing through the same engine.

Open the live demo
shopee.sg/product/slim-coffee-x…
Product image

Miracle Slim Coffee X — lose 5kg in 7 days

GlowMallSG · joined 3 weeks ago
S$9.90 S$89.00
RED — do not buyAccenture Verify
  • Rapid weight-loss + “natural” claims match sibutramine-adulterated products.
  • Priced ~89% below usual retail, a counterfeit signal.
  • New seller, no established Singapore presence.
Assessed risk index 91 / 100 · report to HSA
NATIVE Share-to-check
TikTok Shop · live◉ 2.4k watching
Share sheet
Accenture VerifyCopy · WhatsApp · More
REDReviewed by Accenture Verify
No shareable URL in a livestream? The screenshot is read by the vision model, the same determination, one tap from inside the app.
LIVE DEMO Try the engine

See it work.

The same three surfaces, in-page here on the deck. Score a listing in the portal, watch the classification land in-page, then follow a livestream from Share sheet to verdict.

Open the live demo

The real portal. Seeded HSA cases render instantly for a reliable live demo; any free-text you type is scored by /api/verify for real.

Paste a URL and we read the product name from it; the extension reads the full live page on-device.
Drop an image here, or click to choose
Read by the vision model for in-image claims. A listing fact, not shopper data.
No inspection run yet
Enter a listing on the left, or try one of the sample cases above. Known HSA-alerted products return a certain RED; unknown products return an explainable risk score.
REC. 03 How we built it

Forward-deployed, sovereign, and shippable.

Built the way Accenture's SEA FDE teams ship: agentic tooling for velocity, a fullstack Next.js core for a single deployable, and inference that never leaves the region. The stack a regulator can operate, managed, auditable, and self-hostable when scale demands it.

Data residency in-regionExplainable by designFeatures, not pages
ApplicationNext.js · fullstack
RuntimeECS Fargate · ALB
Data & vectorsRDS Postgres · pgvector
Reasoning & visionBedrock Claude · ap-southeast-1
EmbeddingsCohere Embed · Bedrock
ClientMV3 extension · on-device
The pattern travels

One instrument for product safety, reusable across the whole of government.

The blacklist, the graph, and the share-to-check flow aren't specific to health products. The same engine extends to food safety, cosmetics, and beyond, an HSA sibling to ScamShield, and a template for every agency guarding a marketplace.

120k+
ScamShield precedent, entities blocked
3
consumer surfaces — web · extension · share
1
engine, many agencies