Lantern

A personal AI that knows what you own and watches for safety recalls. Tell Lantern about your car, car seat, crib or space heater. Every day it checks the official U.S. recall lists, tells you in plain words what's wrong and how serious it is, and helps you get the free fix. Everything it remembers is a plain file you can see, edit, delete and download.

Recall apps tell you a recall exists. Lantern tells you what it means for you, and helps you get it fixed.

Built for the Nebius x NVIDIA Global AI Hackathon. All AI runs on NVIDIA Nemotron models through Nebius Token Factory.

Try it: https://lantern-976584717995.us-central1.run.app. Tap See a demo household (made-up people, real recall data), or type your own car. No sign-up. On a phone, add it to your Home Screen to use it like an app.

Built in public. Lantern pivoted from a gaming assistant to a recall watch on 24 Oct 2026; the earlier code is in the git history. Backlog: GitHub Project.

Why

• About 1 in 5 vehicles on U.S. roads (55–58 million) has a safety recall that hasn't been fixed (CARFAX, via ConsumerAffairs). The repair is free, but the owner often never heard about it: the car was bought second-hand, or the letter went to an old address.
• For household products, owner response to recalls averages under 10%. When a recall is only announced in a press release, about 6% of products get fixed; when owners are told directly, about 50% (CPSC figures, via Registria).

Telling owners directly makes them roughly 8× more likely to act. A personal AI that knows what you own is that direct notice.

What it does

	Feature
1	Your things, as visible memory. Cars (by VIN, or year, make and model) and household products, one Markdown file each. Say or type everything at once in the What do you own? box ("our 2018 Civic, a Graco stroller and a crib"): Nemotron Nano picks out each car and product. Also saved from chat, or added one at a time in the app.
2	Build the list from receipts. Nobody types an inventory of their home, but most people have order emails. In My stuff, Add from receipts or order emails: paste order confirmations, upload an order-history file (a CSV keeps only its product-name column; HTML and .eml emails keep only their text, all in your browser), or take a photo of a paper receipt (read by Nemotron Super VL, which copies the item lines and skips prices and card details). Nemotron Super picks out the things that can be recalled (baby gear, toys, appliances, heaters, chargers, car seats…) and skips food, medicine, books and gift cards; up to 25 at once. A brand, model number or product the text never mentions is dropped. Or tap the camera in any "What do you own?" box and photograph the product's label, its box or a receipt: the photo's text is copied, each line is read the same way, and the result goes into the box for you to check before anything is added.
3	Watch things for someone you look after. "My mom drives a 2012 Toyota Camry", typed or said, or "Whose is it?" in the add forms. It gets its own file with owner: Mom, and Lantern says "Mom's 2012 Toyota Camry" everywhere: on Home, in alerts and check-ins, and in chat ("Is my mom's car safe to drive?"). Tell Mom shares the recall as a message written to her ("Hi Mom, Lantern found a safety recall that may affect your 2012 Toyota Camry…"). The person who will act on a recall is the one who hears about it. The demo household's Tucson belongs to Mom.
4	Daily recall watch. Every day Lantern checks NHTSA (vehicles, child car seats and tires) and the CPSC (household products) for everything you own. Home shows what needs attention in plain words: "Don't drive the 2006 Ford Ranger until it's repaired" comes first. Mark each recall fixed or not mine.
5	Is mine affected? A recall covers models; you own one unit. Tap Is mine affected? and Lantern says where to look (from the notice: "Model BYL001 on a label inside the mattress cover"). Type what your label says, or take a photo of it: NVIDIA Nemotron Super VL reads the label, and Nemotron Super compares it with the official notice and answers , or , quoting the sentence of the notice that decides it. "Not included" is only ever said with a quote that appears word for word in the notice. For cars, it walks you through NHTSA's own VIN check (with your VIN ready to copy) and records what it showed. The result is saved in the item's file; you still close the recall yourself.

What Lantern can't know: NHTSA lists recalls by make, model and year. Whether your car's recall has already been repaired is only shown on nhtsa.gov/recalls by VIN, which has no public API. Lantern says "your model is covered by this recall", links that official check, and never claims a specific car is fixed or unfixed.

Data sources

All free and public, with no API key needed (checked live on 24 Oct 2026):

Source	Used for
NHTSA Recalls API (api.nhtsa.gov/recalls/recallsByVehicle)	Vehicle recalls by make, model and year, including "park it" and "park outside" flags
NHTSA vPIC (vpic.nhtsa.dot.gov/api)	Decoding a VIN into make, model and year
NHTSA Complaints API (api.nhtsa.gov/complaints/complaintsByVehicle)	Owners' safety complaints by make, model and year (with crash, fire, injury and death flags), for "What other owners report". Shared per model and refreshed weekly
NHTSA recall dataset (data.transportation.gov/resource/6axg-epim.json)	Child car seat and tire recalls (NHTSA equipment recalls aren't in the vehicle API): the last 10 years, fetched once a day for everyone and matched like household products. Includes NHTSA's completion rate ("24% of these have been fixed so far")
CPSC Recalls API (saferproducts.gov/RestWebServices/Recall)	Household product recalls. Lantern fetches the recent list once a day for everyone and matches it against each person's things locally
Tavily	Reading the makers' own pages (see below), and finding authorized dealers near the user for "Book the free fix"

How Lantern uses Tavily

Official recall notices say that there is a free remedy. The maker's own pages say how to get it, and answer everyday questions about the things you own. Lantern reads those pages with Tavily (app/maker_pages.py):

• How to get the fix. For household products, the CPSC notice names the maker's recall page in its contact details (1,272 of 1,289 notices since 2023 had a web address). Lantern reads it, plus the CPSC's own page, with Tavily Extract (a query about the remedy, the 3 most relevant passages per page). For cars, NHTSA gives no link, so a Tavily Search restricted with include_domains to the carmaker's sites and nhtsa.gov (which hosts the owner letters) finds the page first. Nemotron Super then writes numbered steps using only what the pages say, with the sources linked. Each recall's steps are shared by everyone and kept for 14 days.
• Ask about your stuff. A question about one of your things is searched only on that maker's sites plus nhtsa.gov or cpsc.gov (include_domains), then the best two pages are read with Extract, and the answer cites them. A product's maker site is found once with one search and saved in that product's file as sources, where you can see and change it. If the maker's sites have nothing, Lantern searches the wider web and says so.
• Dealers. One basic search finds authorized dealers near the city or ZIP you type.
• Listings. "Check before you buy" reads a listing link you paste with Tavily Extract (one page). Sites that need a login (Facebook Marketplace, Nextdoor) are not fetched; Lantern asks you to type what's for sale instead.
• Budget. About 3 credits for a new remedy or question (search 1, advanced extract 2); answers are cached, and the whole site has a daily Tavily credit cap. Web pages are treated as untrusted data: the prompts tell the model never to follow instructions inside them.

How we use NVIDIA Nemotron and Nebius Token Factory

Every LLM call goes to the Nebius Token Factory OpenAI-compatible API. Lantern routes each job to the smallest model that does it well:

Model	Used for
Nemotron 3 Nano (MODEL_FAST), thinking off	Reading each message first: pulling out the things you say you own and deciding whether to check recalls; confirming that a CPSC recall really matches a product you own; small talk
Nemotron 3 Super (MODEL_BALANCED), thinking off	Writing answers from the official recall data it was given; rewriting each recall notice in plain English (once per recall, shared by everyone, cached); for "Is mine affected?", working out from each notice how to tell recalled units apart (shared per recall) and comparing your label with it; a plain title and sentence for each group of owner complaints (shared per model, weekly); picking out the recallable things in pasted receipts and order emails
Nemotron 3 Ultra (MODEL_REASONING)	Only when you tick "Take more time to think" in Ask (capped per person and site-wide)
Nemotron 3.5 Super VL (MODEL_VISION), thinking off	Reading the photos you take: a product label for "Is mine affected?" (model numbers, dates, lot codes), a paper receipt for "Add from receipts" (item lines only, no prices or card details), or whatever you photograph with the camera in the "What do you own?" box. Photos are sent once and never stored or cached

Reasoning is switched per call with chat_template_kwargs.enable_thinking. Every call is logged with its model and token counts (data/logs/llm_usage.jsonl).

Hosting: the web app runs on Google Cloud Run. Cloud Run only serves the app; all inference runs on Nebius Token Factory.

Setup

Requires Python 3.11+.

python -m venv .venv
# Windows: .venv\Scripts\activate    macOS/Linux: source .venv/bin/activate
pip install -e ".[dev]"
cp .env.example .env   # then fill in your keys and model IDs

Check the Token Factory connection, then run the app and open http://127.0.0.1:8000:

python scripts/smoke_llm.py
uvicorn app.main:create_app --factory --reload
# or: docker build -t lantern . && docker run --env-file .env -p 8000:8000 lantern

Run the tests (mocked, no tokens or credits spent). Live tests are opt-in:

pytest
RUN_LIVE_TESTS=1 pytest tests/test_live.py

Try the recall clients against the live government APIs:

python -c "from app.tools.recalls import RecallClient; c=RecallClient(); v=c.decode_vin('1HGCM82633A004352'); print(v, len(c.vehicle_recalls(v.make, v.model, v.year)), 'recalls')"

Turn on push notifications locally by generating a key: python -m app.push, then put the output in .env as VAPID_PRIVATE_KEY.

Deployment (Cloud Run, Cloud Storage for user data, Secret Manager, a daily Cloud Scheduler job): see deploy/README.md.

Accounts and privacy

• Try it as a guest: one click, no sign-up. Each guest gets their own isolated space, deleted automatically after 7 days, unless they tap Keep my list to turn it into a free account (username and password only).
• Accounts: username and password (hashed with scrypt). No email address and no tracking.
• Isolation: every user's data lives only in data/users/<id>/, and tests/test_isolation.py proves users can't see each other's data.
• Visible memory: everything Lantern remembers is a Markdown file you can view, edit, delete and download as a zip.
• Ask before acting: Lantern never sends a message or books anything. Booking messages open in your own email, text or WhatsApp app for you to send.
• Only you mark recalls fixed: the assistant has no tool for it, so a web page or a confused model can't hide a recall from you.
• Where your data goes: messages go to Nebius Token Factory to generate replies; searches go to Tavily (for dealers, only your car's make and the city or ZIP you type; for maker pages, the thing's name and your question, never your VIN; for "check before you buy", the listing link you paste); recall lookups send only a make, model and year (or a VIN, to decode it) to NHTSA. VINs are never put in model prompts.
• Voice uses your browser's built-in speech recognition and speech synthesis. In Chrome, speech recognition is done by Google's servers; Lantern says so the first time you tap the mic. Edge uses Microsoft's online speech service (Windows can switch it off: Settings › Privacy & security › Speech); Brave and Firefox have no voice typing. Lantern names the problem when voice can't start. Typing works everywhere.
• Receipts and order emails you paste or upload are read in your browser first (a CSV keeps only product names), then sent to Nebius Token Factory once to pick out the products; only the products are saved, as files in your list.
• Label photos (for "Is mine affected?") are shrunk in your browser, sent once to Nebius Token Factory for Nemotron Super VL to read, and never stored. What it read is shown to you, and saved in that item's file only as a line of text with the verdict.
• Push notifications go through your browser's push service (Google, Mozilla, Apple or Microsoft). Lantern only sends a short title and line, never your list of things.

Skills you control

Everything Lantern can do is a reusable skill with its own SKILL.md in app/skills/ (agentskills.io-style front matter: name, description, plus metadata for the agent tools it owns, the services it uses and exactly what it sends out of the app). The menu's Skills page lists them, shows each file, and lets you turn any of them off. Off really means off for your account: the skill's chat tool answers that it's turned off, its endpoints refuse, its buttons disappear and its reminders stop. Your choices are saved in your own profile.md (skills_off).

Skill	What it does
Recall watch (always on)	Checks your cars and things against the official U.S. recall lists every day, in plain words
Is mine affected?	Tells you whether your one is in a recall, from its label (typed or photographed) or NHTSA's VIN check
What other owners report	Owners' complaints to NHTSA for your car model, flagging problems no recall covers yet
The maker's own pages	Answers questions and gives the steps to get the free fix, from the maker's own site, cited
Book the free fix	Finds authorized dealers near you and drafts the booking message for you to send
Follow-ups until it's fixed	Checks in by notification until a recall is actually fixed
Add from receipts and photos	Adds things from order emails, order history, or a photo of a receipt or a label
Check before you buy	Checks something secondhand for recalls before you buy it

How the agent works

• A lookup planner (Nano) decides whether a message needs a lookup; the loop then runs the tool itself, so the model can't skip it and guess.
• Every tool call goes through one checkpoint (ToolRegistry) with a timeout and an audit log entry in data/users/<id>/logs/audit.jsonl.
• Guardrails: at most 6 tool calls and 40k tokens per message, plus the daily caps below.
• Answers stream over server-sent events with live status lines.
• Web content is untrusted: the model is told never to follow instructions found in fetched pages.

Cost and safety limits

Lantern is a public demo on the developer's own accounts, so every way of spending money has a daily cap. All reset at midnight UTC and are set in .env (0 = no cap):

What	Cap
Model spending, all of it (chat, photos, the daily job)	Each call is priced from Token Factory's list prices in the model router; the site stops calling models for the day at SITE_MODEL_DOLLARS_PER_DAY ($2 on the hosted demo). The recall lists keep working
Messages (anything that calls a model)	30 per guest, 100 per signed-in user, 100 per network (so one person making many guests can't use up everyone's day), 600 site-wide on the hosted demo
"Take more time to think" (Ultra)	10 per guest, 20 per signed-in user, 40 site-wide
Tavily credits	150 site-wide (cached answers are free)
New guests and accounts	30 guests and 5 accounts per network, 300 guests and 100 accounts site-wide
Sign-in attempts	5 failures per username per 10 minutes, 100 attempts per network
Things per person	50 cars and 50 products
Request size	9 MB (a shrunk photo); larger requests are refused

Hosting has its own backstop: deploy/killswitch/ is a Cloud Run function on a Google Cloud budget that takes the app off the public internet if the month's hosting bill reaches $5, so a flood of traffic can't run up a bill. A security review on 27 Oct 2026 also added security headers (a content security policy, no framing), and stopped one person's edited sources from changing the fix steps everyone sees.

Future work

• More recall sources: food and drugs (FDA, USDA), and barcode scanning.
• Signing out everywhere (sessions currently last up to 14 days).
• Parked from earlier versions, in the git history: an animated companion (Glim), a desktop app, and sandboxed helper skills.

Track tools we evaluated

NVIDIA OpenShell, NemoClaw, Hermes Agent and Nebius Serverless were each checked against their official docs: see docs/track-tools-evaluation.md. We kept our own agent loop (multi-user isolation and per-fact visible memory) and borrowed Hermes's memory actions (add, replace, remove).

License

MIT

The fonts in app/web/fonts/ (Geist and Instrument Serif) are under the SIL Open Font License 1.1; their licence files are next to them. Lantern serves them itself, so no third-party font service is contacted.