Prakira is a small tool that turns the next three days of heat, rain, air quality and UV into a few actions for one household, and shows the numbers behind every action, taken straight from the data. This page explains why it was made and which research shaped each choice.
Open weather and air-quality data is now free and very good. A phone can tell you that the heat index will reach 37 °C and the air quality index 204. What it does not tell you is what that means for a grandmother, a toddler, or someone who works on a roof all day.
Language models can write that kind of advice in seconds. They can also state a number that was never in the forecast, in the same confident tone. For a tool about safety, that is the wrong trade.
So the question behind Prakira was narrow: can an AI write the advice without ever being trusted with a number?
These three figures are model forecasts for one grid cell on one day, shown as an example of what the tool reads. They are not measurements and not a trend.
While building it we searched the research literature (36 Scopus queries, 38 references kept). Each card below is one finding and what Prakira does because of it. We read the abstracts of these papers, not always the full text.
Public warning systems send the same message to everyone. People want to know how it affects them now and what they can do about it.
Of 31 heat-health messages written by ChatGPT-4o, most were accurate, none met a grade 8 reading level, and few placed heat in the context of climate change.
Weather-warning experts saw tailored and multilingual warnings as the promising use of AI. Their concerns were over-reliance, less human oversight, and losing a single authoritative voice.
Work on language models for flood communication names hallucination as a key risk, and reports that grounding the model in data cut factual inconsistency by more than 75%.
In a weather-decision experiment, very high and very low false-alarm rates led to worse decisions. Showing the uncertainty of the forecast improved both compliance and decision quality.
A review of heat-health warning systems found most rely on temperature alone, and recommends humidity-aware heat-stress indices, personal factors, and digital tools for targeted warnings.
Being clear about limits is part of the design.
We did not find this combination in our literature search. That is not proof that it is new: we searched one database and did not search preprints.
Prakira was built in October 2026 for ForgeHacks 2026 (track: AI + Climate) by Rofi, a student in Indonesia, with AI coding assistance. The name comes from the Indonesian word prakiraan, forecast.
Your place, household and language are remembered only in your own browser, never on the server, and you can clear them with one click. If you use "Describe your home", that sentence is sent to the AI provider to be read and is not stored by Prakira.
It runs on open data from Open-Meteo and the Copernicus Atmosphere Monitoring Service, and a small open-weight language model from Mistral. It has no dependencies beyond Node.js, and the whole method is public.
The Protective Action Decision Model (Lindell and Perry, 2012) describes how people respond to warnings: they must receive, notice and understand the message, then judge the threat, the action, and the source before they act.
Each briefing item follows that order: what is forecast, for whom it matters, one thing to do, and where the number came from.
Only papers cited on this page. The full review, with all 38 references and the search queries, is in the repository.