Tueanphai: AI flood-risk intelligence for Thailand

01 · Demo video

The whole demo in about five minutes

One recording through every role on the replayed flood day: the duty officer's briefing and questions, a watch that raises an alert, a public message drafted, checked and approved by two people, Max's phone, and the response from incident to SITREP.

02 · How to use

How do I use it?

Five steps take you through the replayed flood day, from the first briefing to a message on a resident's phone.

• Sign in with a roleStart as Duty officer, or look around as a Guest with no passcode. A short guide opens on your first visit; the Guide button at the top replays it.
• Read the AI briefingThe Console opens with the briefing in the assistant panel: what is high risk and why, rivers above their usual flow, what changed since yesterday and what needs a decision. The map colours every area by risk.
• Ask in your own wordsType in the box under the assistant, in Thai or English, or press Ctrl K (⌘K on a Mac). Click an area on the map, or Ask why in a table, to see what drives its score.
• Act on an alertOn Alerts, acknowledge an alert and press Draft message: the AI writes it in Thai and English and the checks compare it with the facts. Submit it; a Supervisor approves and issues it as an exercise, and Max's phone shows what a resident would get.
• Follow the responseOpen an incident from the alert: tasks go to the district with the nearest pumps and boats, the district official reports from the field and an officer verifies the report. Draft a SITREP or a shift handover when you need one.

03 · What you see

What would the user see?

Six pages, each for a part of the work. The top bar shows only the pages your role may open.

• ConsoleThe map and the AI assistant. The briefing opens by itself; ask questions in the box or with Ctrl K. Guests read the briefing and the map only.For: Every role, guests too
• AlertsThe alert inbox, most urgent first: acknowledge, draft the public message, the response plan or an incident. Supervisors approve messages here. The side shows your notices and the active watches.For: Duty officer · Supervisor · District official · Executive · Admin
• IncidentsThe response at a glance: open alerts, incidents and their stage, tasks, field reports to verify and messages waiting. Draft the SITREP and the shift handover here.For: Duty officer · Supervisor · Executive · Admin
• FieldMade for a phone: your district's tasks, and a form to send a field report with an optional photo and your location.For: District official · Duty officer
• Max's phoneMax is a fictional resident in the alerted area. His phone shows the exercise messages a supervisor issued, as cell broadcast, LINE and SMS. A simulation: nothing is sent.For: Duty officer · Supervisor · District official · Executive · Admin
• Audit logEvery AI and human action: who, what, the AI model and the result. Each row's hash covers the row before it, so a changed row shows.For: Duty officer · Supervisor · Executive · Admin

Who does what

Six demo roles, each with its own part. Pick one to try it.

• Duty officer DO-01Briefing, questions, alerts, message drafts and incidents.Try as Duty officer →
• Supervisor SV-01Approves others' messages and SITREPs; can pause all watches.Try as Supervisor →
• District official DI-01Area alerts, district tasks and field reports from a phone.Try as District official →
• Executive EX-01Reads the briefing and follows the response.Try as Executive →
• Admin AD-01Briefing and the audit log.Try as Admin →
• Guest Briefing and map only. No passcode.Try as Guest →

04 · Reading the screen

How do I read the screen?

The labels and colours you will meet, in plain words.

• Risk colours and the 0–100 scoreHigh (▲) from 60, Medium (◆) from 35, Low (●) below. Score = hazard (how likely water is) × exposure (how many people, plus hospitals, are in the way). Hazard weighs the river against its own usual flows (40%), rain expected in the next 72 h (30%), rain in the last 72 h (15%) and new water seen by radar (15%); a part with no data is left out, never counted as zero. A rule of thumb for decision support, not a flood forecast.
• 2-, 5- and 20-year flowThe flow a river point reaches on average once in 2, 5 or 20 years: "above its 2-year flow" means higher than usual for that place. ◉ on the map marks the 20 river points. These are model values (GloFAS), not gauge readings.
• Alert levels and priorityWatch, Warning and Severe are raised by rules in code when a watch's threshold is crossed, never by the AI; the levels are illustrative until an agency gives its own. Priority sorts the inbox (level × exposure × confidence × freshness); open Evidence to see why an alert was raised.
• EscalationA warning nobody acknowledges within 15 minutes goes to the supervisor; a severe one after 5 minutes, to the supervisor and the executive. It runs on today's clock, so the replay shows 2021 data next to today's times.
• The replayed dayThe demo replays 27 September 2021, 08:00 Bangkok, a real flood on the Chao Phraya, on public data. Its 72-hour rain and 7-day river outlook are what then happened (hindsight), not forecasts made that morning.
• AI labelsAI-generated marks text the AI wrote from the tools' numbers: useful, but check it. AI offline means no AI model is connected, so answers come from fixed templates over the same tool results.
• Exercise messagesEvery public message says "Exercise — not an official warning" and has an [ISSUER] placeholder where the agency's name would go. LINE, SMS and cell broadcast are previews only. The person who drafts or edits a message can never approve it.
• Field reportsThe AI reads a district official's report into fields: water depth, rising or falling, needs. It counts as evidence for an alert only after an officer verifies it, and never by the person who sent it.

05 · The issue

What is the current issue? (hypotheses to test with agencies)

• Rain, river and satellite flood data sit in separate systems; joining them is done by hand.
• Warnings come late or are too general; who and what is in the way is worked out under pressure.
• Thresholds exist on paper, but watching many stations and areas at once is hard.
• Alert texts and situation reports are written from scratch each time, in Thai, against the clock.
• A broken or old sensor reading can raise a false alarm, or hide a real one.

06 · How AI helps

How can AI help?

The same five steps run through every part of the demo. The numbers come from code; the AI explains them and drafts; a person decides.

• Data comes inRiver flow outlook, rain expected and rain that fell, water seen by radar, and the people, hospitals and schools in each area.
• Code computesRisk scores, rivers above their usual flow, readings that look wrong, and watches whose threshold is crossed.
• AI explains and draftsThe briefing, answers to questions, and drafts of public messages, SITREPs and shift handovers, in Thai or English, from those numbers.
• A person decidesAcknowledge, approve, verify, open an incident: always a named role. A public message needs two people.
• Everything is recordedEvery AI and human action goes into the audit log, where a changed row breaks the hash chain.

07 · Data and limits

Data and known limits

Public and open data, plus labelled synthetic resources. Thai agency feeds come with a pilot (see Next steps).

Known limits (shown, not hidden)

• River flows are a model (GloFAS): on the dam-regulated lower Chao Phraya it ran about 1.5× above the flow a station measured that day.
• Radar water is detected water, not damage: it includes fields flooded on purpose (retention areas) and paddies.
• Backtest on the replayed day: the score from what was known that morning did not point to where new water appeared better than chance (AUC 0.44–0.59; 0.5 is chance). Radar water seen now is what drives the high scores. One flood, one basin: a check of direction, not proof.
• Teams, pumps, boats and shelter capacities are made up for the demo; the temples and schools used as shelters are real places.

08 · Next steps

What could be developed next?