Seven steps, from checking a listing to the API for AI agents. Every answer cites its source.
Three ways in. A test shop page shows the answer on the listing itself; paste a listing to watch eight agents work; or just ask Eddy.
A real Chrome with the SafeShelf extension. It opens on a real Amazon.in listing, the Janod activity table recalled in the US and Australia (if Amazon holds the page back from our cloud browser, the next tab has a test-shop copy), and our test shops, where all eight test products are real pages; any Amazon or Flipkart link works too.
Eight shop pages where we already know the right answer, including a safe product that must stay quiet.
The right answer is known for each, and each works both ways: an instant test page, or a real shop page in the real browser.
A shared cloud browser: don't sign in to your own shopping accounts here. It closes by itself after 5 minutes.
Copy a listing from any shop and watch eight agents check it, one by one, on the live API.
The same chat shoppers use on the website. Ask in your own words, paste a link or drop a screenshot.
The real SafeShelf chat on the live API, the same one shoppers use on the website. It remembers the conversation and shows its sources under every answer.
Open full screen ↗Recall notices are English web pages. The people they miss shop on a phone, often in another language.
of banned or recalled products inspected were still for sale online.
OECD product safety sweepproducts without an ISI mark, or with a fake one, seized in a single raid on e-commerce warehouses in Delhi.
BIS via PIB, 27 Mar 2025recall databases in India. Orders and raids sit in English press releases; we gathered 36 of them, 2021–2026.
Our dataMade for people who would rather not read small print. Talk to it with the camera on, or ask with a photo of the label.
Made for people who would rather not read small print: older shoppers, people who read little, and people who can't see the screen.
Forward a shop link or a label photo and get a text and a voice note back. The chat here runs on the live API.
How the SafeShelf WhatsApp number works: forward a shop link or a photo of the label, and get the answer as a text and a voice note, in English, Hindi or 24 other languages. This chat runs on the live API. The WhatsApp number is coming soon.
Loading the chat…
For shoppers who live in shopping apps, not browsers. Share a product from Amazon or Flipkart to SafeShelf, and Eddy checks it with the same live API.
Version 0.2.0, installed on our test phones. The source is in the repository.
See the Android sourcePlanned next"Still on sale" report for IndiaCheck before you gift, for DiwaliFree for shoppers, NGOs and researchersConsumer groups we plan to approach; none is a partner yet
Most recalls arrive after the product is home. SafeShelf remembers what you bought and tells you when that changes.
Saved products are rechecked for new recalls
Kelvo CX-65 chargeramazon.in · saved 3 OctCAREFUL
NEWDERY power bankflipkart.com · saved 9 OctRECALLED
Electric kettleflipkart.com · saved 12 OctALL CLEAR
Typed in, or read from a receipt photo
Return windowends 21 Oct 20263 DAYS
Warrantyends 12 Mar 2027OK
Receiptphoto · Flipkart orderREAD
Answers only from the manual, with page numbers
Can I charge it overnight?p. 14 · SafetyNO
What do the 3 lights mean?p. 6 · Statusp. 6
A one-time code, read-only, either side can stop
Ammasees your shelf · read-onlySHARED
Invite codevalid for 24 hours4F7K-2Q
Built and tested on Vertex AI; the household tools go live with the next API deploy. Until then Eddy answers these from the same rules.
Heat, swelling, a burning smell. Safe steps first, then the paperwork: refund, complaint, report, disposal.
Never “test it”; emergency numbers per country
1 · Unplug it, if you can do so safelydon't touch it if it is swellingNOW
2 · Move it onto a hard surfaceaway from bedding and curtainsNOW
3 · Fire or smoke: call 112India; numbers for every countryCALL
The Report draft agent fills it in from the case
To: the sellerrefund request, case attachedDRAFT
To: the consumer helplinecomplaint, with the evidenceDRAFT
From the extension; a person reviews it first
Heat report · CX-65joins 9 reports on all versionsUNDER REVIEW
Your detailsnever stored, only a hashed referencePRIVATE
Official battery and e-waste collection
Lithium batterynot in household wasteE-WASTE
Whereauthorised collection points, by countryFIND
The same engine, at catalog scale. Every finding comes with its case file, so a person can act on it.
Loading the catalog scan…
Loading the evidence pack…
Shopping assistants ask SafeShelf the same questions: ten read-only tools over MCP, a streaming REST API and an A2A agent card.
Loading the API playground…
Gemini does the reading, searching and talking. Plain code makes the final call, so it cannot invent a recall.
Two questions about every listing, then one of four answers, each with the evidence behind it.
Code picks the answer. Plain rules turn the evidence into a colour; Gemini only explains it, in your language.
Exact model only. A look-alike's recall is never blamed on the wrong product.
Tested on 25 real recalls and 40 products that were never recalled, against Gemini alone and keyword rules.
We attacked our own code: fake regulator pages, look-alike models, hidden instructions. Here is what it decided.
Why we built it, who it helps, what it runs on, and where it still falls short.
Recall notices exist. Shoppers just never see them at the moment they buy.
Product recalls are published as English web pages on regulators' sites, one country at a time. A power bank recalled in the US for fire can still be listed on an Indian or Australian marketplace, under the same model number or a new brand name. SafeShelf checks a listing at the point of purchase and answers two questions with evidence:
A shelf is where you keep what you bought; SafeShelf watches it before and after you buy. Eddy is the guide who talks you through the answer, in chat, by voice or on WhatsApp. Eddy explains; code decides.
India, Japan, Australia, Singapore and Southeast Asia buy the same mass-market electronics as the US, but most published recall data is American: India has 36 notices on file against 2,537 from the US. SafeShelf loads 8,689 records from 10 regulators and carries a US recall across the border as "recalled abroad, same model", never as "recalled here".
A chatbot can invent a recall or miss one. In SafeShelf, Gemini reads, searches and explains, but plain code sets the colour from the evidence, and only an exact model match inherits a recall. When the evidence is thin it says "Can't tell yet".
On 5 May 2026 a power bank ignited in a seat pocket after IndiGo flight 6E-108 landed at Chandigarh; about 200 passengers were evacuated and at least 6 were hurt. No brand or model was named.The Times of India, citing the DGCA
Lithium batteries, heaters and children's products cause harm long before a regulator acts. In our early work, a product recalled in one country was recalled in a second a median of 22 days later, and up to 293 days later.
It answers where people already shop and talk: on the listing (Chrome extension), on WhatsApp (text and voice notes), in an Android app and on the web. Answers can be read aloud in 26 Asia-Pacific languages, and it works with screen readers and the keyboard. See step 2.
My Shelf keeps watching saved products and shows when a status changes. If something goes wrong, SafeShelf gives safe steps first, then a refund request and a complaint to the right regulator (India NCH 1915, ACCC, Japan's CAA and others), already filled in with the official record. It drafts them; it never sends anything on the shopper's behalf.
Yes. The same engine scans a marketplace catalog and says what to do with each finding (stop selling, quarantine, notify buyers): on 500 test listings it found 49 of 50 planted problems with no wrong findings. Regulators get an evidence pack of products recalled abroad but still sold locally, with a case file per finding. See step 5.
Google Cloud, end to end. One engine serves every channel.
A live check takes a median of 6.6 seconds.
Regulators' own feeds and pages, plus 29,379 real buyer reviews and an 8,470-record Australian energy registry. Sources that block automated access are shown as "couldn't be checked", never as clear.
Team Hashes, two builders in a week for the AI Builder Cup 2026, Retail & Commerce: Solomon, the brain (engine, data, API, MCP, WhatsApp, benchmarks) and Seshu, the face (extension, ADK agents, web, voice). All the code is on GitHub.
On 25 real US recalls and 40 never-recalled products, reading only reviews written before each recall, SafeShelf flagged 6 recalls with 2 false alarms. Keyword rules flagged more (11) but raised 11 false alarms. On the recalls we did not tune on, it caught 2 of 9 with no false alarms. The sample is small, and every number is in Results.
We attacked it ourselves, with Gemini writing the attacks: fake regulator pages, look-alike model numbers, hidden instructions aimed at AI tools. The red team found two real holes, which we fixed: lightly reworded reposts, and fake-shop notices "for AI shopping assistants" (now 13 of 13 caught, no false alarms on 27,822 real listings). See step 7.
No. Green means no problems were found in what was checked, and every answer lists what was checked and what could not be. Verdicts are evidence for shoppers, not legal findings; AI findings marked "under review" are not regulatory determinations.
Real: recalls, reviews and registries. Fictional: Kelvo, SafeVolt, VoltNova, Zorvana and MegaDealz24, so the demo can show every answer without accusing a real seller. Cross-border "still on sale" matches are leads a person has not yet confirmed, so we do not count them as findings.
Shoppers use it free. Marketplaces, insurers, payment providers and regulators would pay for catalog scans, purchase risk flags and the recall-check API. Tenant keys, webhooks and usage metering are already built.
The web page and the API work today. The Chrome extension, WhatsApp number and Android app are coming soon.
EddySafeShelf chat
