Try it: Check what you own | How it was measured | The business | Three recalls caught before the databases | A replayed watch | Code

Tagout: type or photograph what you own; a Nemotron 3 Ultra agent on Nebius Token Factory checks CPSC, NHTSA, FDA, USDA and the web (Tavily), and hangs a tag on each thing: Do not use, Check the label, or Checked.

Why it matters. CPSC's own staff found that recalls sent straight to owners got 50% of them to act; recalls announced only by press release got 6% (CPSC Recall Effectiveness Workshop, closed cases FY2013–2016). Tagout is that direct notice for the people who own the things: households (about 187 of CPSC's 498 recalls of 2026 name a child's product), landlords (California's AB 628 now makes them fix a recalled stove or fridge within 30 days of notice) and resale sellers (counted Oct 29, no model).

The numbers that matter

  • On 66 cases it had never seen, in three held-out sets each run once (the last two pre-registered with public git tags): 42 of 46 recalled items flagged and 3 false alarms on 20 clean items. It fails safe: since set 1's fixes, 0 of 15 recalled items were cleared green across 27 pre-registered cases.
  • Tavily, measured: on the two held-out sets with a Tavily-off arm, Tagout with Tavily off cleared every recall no database had yet (10 of 10, 2 of 2); with it on, it flagged 8 of 10 and 2 of 2.
  • About 4.4¢ an item on the 105 tuned cases (2.6¢ of Nemotron, 1.7¢ of Tavily); 4.7-5.5¢ on the held-out sets. Every answer, misses included.

Test it in one minute

Inspiration

Each recall lives in its own database, in its own words, and the detail that decides whether yours is covered (a model, a lot, a date window) is buried in a paragraph. We wanted "is any of my stuff recalled?" to take two minutes, with proof, and without an AI guessing.

What it does

The home page replays a real check: every tool call Nemotron made, the notices read, the code matched and the tag hung.

Nemotron does what rules can't; rules make it provable.

  • Reads what you own. Type it, or photograph the label. Nemotron turns each line into an item with the codes a recall turns on (model, lot, serial, VIN, UPC, NDC); a code that isn't in what you gave is dropped.
  • Searches like an investigator. An agent on NVIDIA Nemotron 3 Ultra calls six tools: NHTSA vehicle and child-seat recalls, CPSC, openFDA, Tavily search and Tavily extract, and always searches the web, because the databases trail the agencies' newsrooms by one to six weeks.
  • Reads every notice's list. Nemotron reads the affected models, lots, UPCs and date windows; every code it returns must be printed in the notice, as whole words.
  • Rules hang the tag: Do not use (your code is listed or inside a printed range), Check the label (the line is recalled, yours isn't confirmed; cars always stop here), Checked (nothing lists it today). With no model answer, nothing is cleared green.
  • Watches. A daily sweep reads each new recall once and matches it against every saved list by the same rules. Alerts reach you today through the watch's feed, in a feed reader or a feed-to-email service; built-in email and push are next. Most days the sweep makes no model call.

How we built it

How Tagout is built: Nemotron on Nebius Token Factory reads the list, runs the search agent and reads each notice; Tavily covers the web; rules decide; a daily sweep watches.

  • Nemotron on Nebius Token Factory, each role measured. Nemotron-3-Ultra-550b-a55b keeps the list reader and the agent (Nano dropped a car's year; Super filed medicines as products, the wrong database) and reads notices; NVIDIA-Nemotron-3-Nano-30B-A3B reads notices in the lite tier (101/105 for 0.23¢ an item); Nemotron-3.5-Super-VL reads label photos. Reasoning is off (enable_thinking: false) where we measured it doesn't help.
  • Tavily: search locked by include_domains to the official recall sites, with time_range for new recalls; extract to read a page in full, including FSIS's PDF product lists, so a UPC is matched with its own lot. FDA company announcements and USDA FSIS recalls are reachable only this way.
  • The daily sweep reads each recall once for everyone (cached in Google Cloud Storage). Oct 29, 06:02 UTC: 493 recalls, 10 new, 0 model calls, 2 Tavily credits, 9.4 s. It runs as a Vercel cron; a Dockerfile for a Nebius Serverless Job is included (not yet run there).
  • App: Next.js 15, streaming each check; GitHub Actions CI and a daily Token Factory canary; 120 unit tests.

Results

Three held-out sets, each run once, and Tavily on and off.

Set (each run once) Tagout right False alarms on clean items Recalled items cleared green
Held out #1: 39 cases written Oct 25, committed before any run 31/39 0/8 4/31 (each fixed since)
Held out #2: 19 cases, pre-registered with tag prereg-2026-10-25 15/19 1/8 0/11
Held out #3: 8 cases, pre-registered with tag prereg-2026-10-30 4/8 2/4 0/4
105 tuned cases (in-sample; the first run was 98/105) 105/105 0/38 0/67

"Recalled items" are the cases whose right answer is Do not use or Check the label, and "flagged" means either tag. A few recalled lines and cars are expected to get Check the label, so they also count where Do not use would be wrong; /proof's "wrongly red" columns include them (0/10, 1/9, 2/4; 0/59 on the 105).

  • The agent against the lite tier (fixed search plan, Nano reader): 50 vs 48 over the three held-out sets; the lite tier won set 3, 5 vs 4. The fixed plan with Ultra readers ran on set 1 only (agent 31, plan 29). Only the agent got h07, h43, g06 and g14, each by a search the plan didn't run (on /proof), so the agent runs the first check and the cheap sweep re-checks.
  • Set 3 ran with Tavily's keyless cap refusing most searches, so it measured the database path. Its misses, and the in-sample fixes since, are on /proof.

Why not a recall search box, a chatbot or an alert app

  • CPSC's database never fills in its model field (empty in all 2,315 recalls since 2020, checked Oct 29); Tagout reads the notice text and verifies every code against it.
  • A model alone invents recalls. Nemotron from memory wrongly tagged 5 things red on the 105, each time citing a recall that doesn't list the item.
  • Alert apps match the code you enter; Tagout adds a held-out error rate, the agency's own sentence beside every red tag, and recalls found on the web first.

Limits

  • U.S. recalls only. Green means nothing listed the item on the day of the check. Cars stop at "Check the label" (no public per-VIN API).
  • Bounded codes are the weak spot: set 2's and set 3's misses were all codes or dates a recall bounds instead of listing. They are fixed since, in-sample; one set-3 unit (letter-only part codes) stays orange where red is right.
  • The demo runs on shared credits with daily caps, so busy days fall back to the lite tier.

Challenges we ran into

  • Codes in parts and in ranges (a lot printed 091026 DL, a serial inside a printed table of ranges), each now matched as printed, with tests; agency quirks such as CPSC's fake error records, openFDA's empty brand fields and FSIS turning scripts away (Tavily reads it).
  • Rate limits: the app never waits on Tavily's keyless cap; it goes on with the databases and says so.

What's next

  • A fourth pre-registered held-out set; Health Canada and EU Safety Gate.
  • Built-in email and push alerts; a property-manager import and a resale-listing API. Households use it free; landlords and resale platforms would pay (/business); no paying customer yet.

The team and the video

Built by Rohan Borade (design and build) and Aatmaja (recall research and testing). The demo video (voice recorded Oct 25, cut Oct 30) shows the replayed check, the Nemotron 3.5 Super VL label read, Tavily catching recalls ahead of the databases, the daily watch alerts and all three held-out sets.

Built With

  • cpsc-api
  • gemma
  • github-actions
  • google-cloud
  • nebius-token-factory
  • nemotron-3-nano
  • nemotron-3-super
  • nemotron-3-ultra
  • nemotron-vl
  • next.js
  • nhtsa-api
  • nvidia-nemotron
  • openfda
  • qwen
  • react
  • tavily
  • typescript
  • vercel
Share this project:

Updates

Submission history