Antidote

Photograph a pesticide bottle and get the chemical, its WHO hazard class and the national guideline's own words on treatment, or a stop when no antidote exists, in under 30 seconds.

Built on Nebius Token Factory with NVIDIA Nemotron 3 Nano, 3 Super, 3 Ultra, 3.5 Lightning and 3.5 Super VL. Best Apps and Agents track. Apache-2.0.

Four days to name the poison. In July 2013, 23 children in Gandaman, Bihar, died after a school lunch was cooked in oil kept in a pesticide container. Police named the chemical, monocrotophos, four days later (Al Jazeera). On a pesticide bottle, the name is printed on the label. Antidote reads it, for about 4 US cents of model calls per card.

Decision support for trained clinicians. Airway, breathing and circulation come first. Every clinical sentence on a card is quoted from a national guideline, with its page. National Poisons Information Centre, AIIMS: 1800-116-117.

Try it in two minutes (no sign-up, no key)

What to try	Link	What you will see
A real label, live	Gramoxone (paraquat): click ingested, then Build the card	Two stop cards first (no antidote; no extra oxygen unless hypoxic), then a card made only of guideline quotes with their pages, an Ultra audit, and every model call with its cost
Two bottles, two treatments	Landing page: switch between Gramoxone and Dursban	The toxicity triangle, WHO class, stop cards and guideline line change with the bottle
The district view	Register	A spraying-cluster alert and a district brief, from demo register rows marked as synthetic
The results	Evals page	0 unsafe cards on 123 test cases; 107 unsourced doses from a general model against 0 from Antidote
Your own hospital	Data page	Add your own protocol and stock list; cards cite your protocol
Your own AI agent	Add https://antidote-api.vercel.app/mcp as a remote MCP server in Claude, Cursor or VS Code	Four tools, no install: list_chemicals, bedside_card, read_guideline_page, surveillance_alerts

Web: https://antidote-bay.vercel.app. API: https://antidote-api.vercel.app (OpenAPI docs at /docs). Demo video: https://youtu.be/S8XQYCcvGmM

Why it matters

In 94 percent of recorded acute pesticide poisonings in India, the chemical went unidentified (PAN India, 2026). Yet every Indian pesticide label must carry the active ingredient, a colour-coded toxicity triangle and an antidote statement. The chemical decides the treatment:

• In an Indian paraquat series, patients who could not name the poison were first treated for organophosphate poisoning, with atropine and oxygen, and oxygen worsens paraquat injury. The diagnosis was made from the bottle, or from photos of it shared on WhatsApp by relatives (Dambal et al., 2021).
• In a Bangladesh hospital series, 63 of 198 patients (31.8 percent) were misdiagnosed from symptoms alone (Verma, Eddleston et al., 2024).
• Case fatality ranges from around 70 percent for paraquat to close to 0 for many newer pesticides; WHO tells families to "take along the bottle" (WHO, 2008).
• Where poison-centre guidance reaches the bedside, patients leave hospital 3.43 days sooner on average (Galvão et al., 2011).
• In Yavatmal in 2017, about 800 cotton farm workers were admitted to hospital after spraying and more than 20 died before an investigation was ordered (Public Eye). Every one of them was in a hospital register.

The full reference list is on the live site at /references.

What it does

1. Photograph the label. Two different vision models read it. The chemical is used only when both reads agree and it appears in the text read.
2. Name the chemical. Antidote matches it to India's pesticide registry, with its WHO hazard class and Indian regulatory status. Tavily checks unknown brands.
3. Stop first when it matters. If no antidote exists (paraquat, for example), a stop card comes first. Deadly mistakes, such as atropine for paraquat, are blocked by fixed rules, not by the AI.
4. Build the bedside card. Every clinical sentence is quoted word for word from the NHSRC poison-management module or the ISCCM 2024 position statement, with its page. The model never writes a dose.
5. Get a second opinion. Nemotron 3 Ultra reads both guidelines in full and audits the card.
6. Refuse to guess. A smudged or torn label gets a guarded card asking the doctor to confirm from the bottle.
7. Fill the paperwork. One click saves the medico-legal register row, pre-filled from the label.
8. Warn the district. Fixed rules flag a spraying cluster (3 cases of one chemical group in 7 days), banned chemicals still reaching patients, and restricted ones. Nemotron 3.5 Lightning writes the district officer's brief.

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗