Agents for Humans: thunAI_UGMDU in Action - From Early Warning to Real-World Help
A walkthrough of thunAI_UGMDU across its three surfaces - the public status page, the coordinator console, and the responder app - showing how a single flood signal moves from early warning to real-world help. It follows one incident end to end: the agent detects and assesses the hazard, warns the community, asks a human before any irreversible action, and coordinates the response through to a verified outcome.
Series: Agents for Humans: Building thunAI_UGMDU - A Neighborhood AI Agent (3 articles)
- 3Agents for Humans: thunAI_UGMDU in Action - From Early Warning to Real-World Help This article
📌 Public Status - Information for Everyone
The first surface is the one that matters most in an emergency: the public status page. It is open to anyone, requires no account, and is designed so a resident can answer a single question at a glance - is my neighbourhood safe right now?
What the page shows:
- Current status - the community's severity band, shown as a single clear label with a plain-language recommended action.
- Why - a factual, one-line reason for the band, built only from readings that actually crossed their danger threshold (never model-generated text).
- Affected areas - the areas involved and how many requests have been reported in each.
- Shelters - open shelters and how many places are still available.
- Current levels vs danger threshold - river level, rainfall rate, and dam release, each shown as the current value against the danger line, so residents see the evidence, not an internal rule dump.
- Reading time - when the data was last updated, so people know how current it is.
Crucially, there is no login, no app to install, and no special knowledge required. Everything on this page is aggregate and non-identifying - no resident's name, contact, or household location ever appears.
The four status bands
The page always reflects one of four escalating states, so the community sees the same clear signal whether things are calm or critical:
✦ NORMAL - no active hazard. Readings are within the safe range; no action needed. The page simply confirms the neighbourhood is safe.
✦ WATCH - conditions are worth monitoring. Something has begun to move; residents are asked to stay aware and keep an eye on updates.
✦ WARNING - a hazard is developing. The recommended action becomes concrete - prepare to move, keep an evacuation kit ready - and affected areas, shelters, and levels-vs-threshold are all shown so people can act early. (This is the state shown in the screenshot: river level 4.5 m against a 6 m danger line, rainfall 38 mm/h against 50 mm/h.)
✦ EVACUATE - the danger threshold has been crossed. The page reflects the most severe state, names the affected areas and open shelters, and the reason line states plainly which readings have exceeded the danger threshold.
The message:
People shouldn't need an account, an app, or special knowledge just to know whether their neighbourhood is safe. thunAI_UGMDU treats clear, verified, public information as the baseline - one trustworthy source of truth that updates in real time as the situation changes, from all clear through to evacuate now.
🦚 Tamil - AI That Speaks the Community's Language
An emergency system is only as useful as the number of people who can actually understand it. In many neighbourhoods, the residents most exposed to a flood are not the ones most comfortable reading English - and in a crisis, a message that has to be translated in someone's head is a message that arrives too late.
That is why thunAI_UGMDU treats regional language as a core capability, not an afterthought. The public status page can be viewed entirely in Tamil, from the heading (சமூக வெள்ள நிலை - "Community Flood Status") to the current status (வெளியேறவும் - "Evacuate"), the recommended action, the affected areas, the shelter list, and the current-levels-versus-threshold table. The same underlying data a coordinator sees in English is presented to a resident in the language they think in.
Language matters most precisely when the situation is stressful. Under pressure, people fall back to their first language; comprehension of a second language drops sharply when fear and urgency are high. A warning that says "prepare to move, keep an evacuation kit ready" is only actionable if it is read in the reader's own language, immediately, without hesitation. Presenting emergency information in Tamil removes a layer of friction at exactly the moment friction is most dangerous.
This carries through the whole workflow, not just the display. Residents can send their requests in Tamil or English - the Intake agent understands both, and code-mixed Tamil-English - and the Alert agent drafts warnings in the local language rather than English-only text. So the community both receives information and interacts with the system in a language it is comfortable with.
The point is not that we added a Tamil translation. The point is a principle:
An emergency system is only useful if the people who need it can understand it - instantly, and in their own words.
That is what makes this part of the "For Humans" story rather than a feature line: accessibility in the local language is what turns early warning into early action for the people who are actually at risk.
☀️🌙 Day & Night - Designed for Real Conditions
Emergencies do not keep office hours. A river rises overnight, an alert arrives at 2 a.m., a responder checks their assignment in the dark on the way to the scene. People do not interact with an emergency system only in bright daylight - and a screen that is hard to read in the conditions someone is actually in is a screen that slows them down when every second matters.
thunAI_UGMDU is built for both. The same interface - resident status page, coordinator console, and responder app - offers a day mode and a night mode, switchable from a single toggle, so the display suits the environment rather than fighting it.
- Day mode uses a light background with high-contrast dark text, comfortable in bright, outdoor, or well-lit conditions.
- Night mode uses a deep navy background with light text and calibrated contrast, so the screen is legible in the dark without the harsh glare of a white page - easier on the eyes at night and less likely to draw unwanted attention in the field.
Critically, the content does not change between modes - only its presentation. The severity band, the recommended action, the affected areas, the shelters, and the levels-versus-threshold table read clearly in both, and the colour-coded status (the red EVACUATE badge, the "above danger" markers) stays high-contrast and unmistakable in either theme.
This matters equally for all three audiences: a resident reading a warning at night, a coordinator monitoring an unfolding incident through the small hours, and a responder acting on an assignment in low light. Usability at any hour is not a cosmetic nicety here - it is part of making sure the information reaches people in a form they can actually use, whenever the emergency happens.
The point is simple: an emergency system has to work in the real conditions people are in - day or night, indoors or out - not just in the conditions we design it in.
👤 Coordinator - From Information to Coordination
The coordinator console is where the human stays in control. thunAI_UGMDU does the heavy lifting - watching conditions, triaging requests, drafting responses - but every consequential decision surfaces here for a person to make.
What the coordinator can actually do:
🎯Review what needs a decision - the Decision Inbox lists each pending action as a clear card: what the agent wants to do, why you're being asked, the stakes, a live countdown to the deadline, and what happens by default if no one responds.
🎯Understand priority at a glance - irreversible or mass-audience actions (an evacuation advisory to ~450 residents, a boat rescue for non-ambulatory residents) are the ones that land in the inbox; routine, safe actions never do.
🎯Approve or decline - one tap approves the action or holds/declines it, keeping the call with a human.
🎯Coordinate response - approving a dispatch triggers a live handshake that creates the responder's assignment instantly.
🎯Monitor incidents and progress - the open-incidents view shows the current band, affected areas, and shelter availability; responder status flows back as it changes.
🎯Resolve and audit - incidents close when resolved, and every step is recorded in a per-incident audit trail the coordinator can review.
The core message: the AI helps understand and prioritise the situation, but the consequential decisions - the ones with real-world, irreversible impact - remain firmly under human control.
🎯Understand priority at a glance - irreversible or mass-audience actions (an evacuation advisory to ~450 residents, a boat rescue for non-ambulatory residents) are the ones that land in the inbox; routine, safe actions never do.
🎯Approve or decline - one tap approves the action or holds/declines it, keeping the call with a human.
🎯Coordinate response - approving a dispatch triggers a live handshake that creates the responder's assignment instantly.
🎯Monitor incidents and progress - the open-incidents view shows the current band, affected areas, and shelter availability; responder status flows back as it changes.
🎯Resolve and audit - incidents close when resolved, and every step is recorded in a per-incident audit trail the coordinator can review.
The core message: the AI helps understand and prioritise the situation, but the consequential decisions - the ones with real-world, irreversible impact - remain firmly under human control.
🚑 Responder - Turning Decisions Into Physical Help
This is the other side of the workflow - where an approved decision becomes real-world action. Coordinator → Responder → help on the ground.
The moment the coordinator approves a dispatch, a live handshake creates the assignment in the responder's app. The responder opens Your assignment and sees everything needed to act, with no back-and-forth:
🔔 The assigned request - a clear "Rescue assignment" card (#REQ-BF4B) with a status badge that starts at Awaiting your acknowledgement.
🔔 Incident details - occupants (2), mobility assistance needed (Yes), medical need (No), and the equipment required (Rescue boat).
🔔 Location and priority - the location reference (Riverside Lane, near the bridge) and an Acknowledge by deadline, so the responder knows where to go and by when.
🔔 Accepting the assignment - Accept takes it on (or Decline to send it back for reassignment).
🔔 Updating status - a clear stepper drives the lifecycle: Assigned → Accepted → On my way → Arrived → Rescue complete, with each stage time-stamped on the timeline as the responder taps through it.
🔔 Completing the response - Rescue complete closes the responder's part; the badge becomes Rescue complete - awaiting coordinator verification, and that status flows straight back to the coordinator console.
The lifecycle is deliberately simple, with friendly labels ("On my way", "Arrived", "Rescue complete") and large tap targets - because the person using it is in the field, often in poor conditions, not at a desk. And because every stage is time-stamped and fed back to the coordinator, the loop closes with evidence: not just that a rescue was dispatched, but that it was accepted, carried out, and completed.
🔔 Incident details - occupants (2), mobility assistance needed (Yes), medical need (No), and the equipment required (Rescue boat).
🔔 Location and priority - the location reference (Riverside Lane, near the bridge) and an Acknowledge by deadline, so the responder knows where to go and by when.
🔔 Accepting the assignment - Accept takes it on (or Decline to send it back for reassignment).
🔔 Updating status - a clear stepper drives the lifecycle: Assigned → Accepted → On my way → Arrived → Rescue complete, with each stage time-stamped on the timeline as the responder taps through it.
🔔 Completing the response - Rescue complete closes the responder's part; the badge becomes Rescue complete - awaiting coordinator verification, and that status flows straight back to the coordinator console.
The lifecycle is deliberately simple, with friendly labels ("On my way", "Arrived", "Rescue complete") and large tap targets - because the person using it is in the field, often in poor conditions, not at a desk. And because every stage is time-stamped and fed back to the coordinator, the loop closes with evidence: not just that a rescue was dispatched, but that it was accepted, carried out, and completed.
The point: thunAI_UGMDU doesn't stop at a decision. It connects that decision to the person who carries it out, and verifies that help actually reached the people who needed it.
🔄 One Complete Journey - Signal → Resolution
Every part of thunAI_UGMDU exists to serve one continuous flow. Here is that flow, told as a single real incident on Ward-7's riverside - using the actual data and screens from the running system.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
Community
│ signal / incident
▼
thunAI
│ understand & assess
▼
Coordinator
│ approve
▼
Responder
│ real-world response
▼
Resolution
│
▼
Community status updated
Signal. Heavy rain over Ward-7. On its scheduled sweep, thunAI_UGMDU ingests the readings - river level 6.1 m and rising, rainfall 55 mm/h, dam release 2,850 m³/s.
Understand & assess. The deterministic Rule Engine bands the hazard EVACUATE; the Monitor opens an incident for the riverside area. The public status page updates for everyone - no login - showing the band, the reason ("River level (6.1 m), Rainfall rate (55.0 mm/h) and Dam release (2850.0 m³/s) have exceeded the danger threshold"), the affected areas, and open shelters.
The agent decides - Act, Ask, or Stay Quiet. This is where the philosophy governs everything:
🔸 STAY QUIET - routine readings within safe range produce no alert. The agent doesn't create noise when nothing needs to happen.
⚠️ ACT - safe, clear steps (publishing the status, drafting the warning in Tamil and English) proceed automatically.
🚨 ASK - the two consequential calls are held for a human. Both land in the coordinator's Decision inbox: "Send an evacuation advisory to Ward-7 riverside residents?" (stakes: ~450 residents) and "Dispatch a boat rescue to 2 residents needing mobility help at Riverside Lane?" (an irreversible commitment of a limited asset). Each card shows why you're being asked, the stakes, a live countdown, and the safe default if no one responds.
Coordinator. The coordinator reviews and approves the boat dispatch - the decision stays with a human.
⚠️ ACT - safe, clear steps (publishing the status, drafting the warning in Tamil and English) proceed automatically.
🚨 ASK - the two consequential calls are held for a human. Both land in the coordinator's Decision inbox: "Send an evacuation advisory to Ward-7 riverside residents?" (stakes: ~450 residents) and "Dispatch a boat rescue to 2 residents needing mobility help at Riverside Lane?" (an irreversible commitment of a limited asset). Each card shows why you're being asked, the stakes, a live countdown, and the safe default if no one responds.
Coordinator. The coordinator reviews and approves the boat dispatch - the decision stays with a human.
Responder. Approval triggers a live handshake: assignment REQ-BF4B appears instantly in the responder's app - Riverside Lane, 2 occupants, mobility assistance, rescue boat. The responder accepts and moves through On my way → Arrived → Rescue complete, each stage time-stamped.
Resolution & community update. The completed rescue flows straight back to the coordinator console - the Active dispatches panel shows Rescue #REQ-BF4B - COMPLETED. The incident closes, the public status updates, and every step from signal to resolution is recorded in an append-only audit ledger.
That is the whole arc - Signal → Understand → Ask → Act → Respond → Verify → Resolution - with one principle running through it: the agent acts when it's clear, asks when the stakes are high, and stays quiet when nothing needs doing. thunAI_UGMDU turns a rising river into help on the ground, and keeps a human in control exactly where the consequences are real.
🚀 Behind the Scenes: Observability, Assistance, and Live Situational Awareness
thunAI_UGMDU is not only about responding to an emergency - it also keeps the system observable and the people involved informed.
Coordinators can review recent agent runs and their execution details - trigger type, start time, outcome, token usage, latency, model, and estimated cost - giving clear visibility into how the AI workflow is performing and what it costs to run.
The coordinator assistant offers a conversational way to ask about incidents, requests, responders, and community-safety guidance. Crucially, it is read-only: it can surface and explain information, but it cannot make changes on its own - consequential actions still flow through the Decision Inbox for human approval.
The incidents view gives the coordinator a live operational picture: the current severity band and evacuation status, affected areas, the count of open requests, assigned responders, and available shelter capacity across the ward - all kept current automatically.
Together, these capabilities give the coordinator both AI assistance and operational visibility, while keeping the important actions firmly under human control. The system doesn't just act — it shows its work, explains itself when asked, and stays accountable.
💰 Cost Breakdown for thunAi
Architecture is almost entirely serverless + pay-per-use
🤖 Why This Is an Agent, Not Just an App
It's worth being precise about what thunAI_UGMDU is, because the distinction is the whole point.
A normal dashboard shows information - it presents data and waits for a person to interpret and act on it.
A chatbot answers questions - it responds to a prompt, then stops.
thunAI_UGMDU does something different. It understands signals → makes decisions → coordinates people → takes appropriate action → verifies the outcome. It runs on its own cadence, reasons over an evolving situation, decides whether each action is safe to take, refers the consequential ones to a human, drives the response through to the field, and confirms that help actually arrived. That full loop - not any single reply - is what makes it an agent.
And the human-centered features are not decoration; each one reinforces exactly that:
🛡️ Tamil → it understands the community, in the language people actually think in.
🛡️ Day / Night → it adapts to the real conditions people are in, at any hour.
🛡️ Public Status → it keeps everyone informed, with no login and one source of truth.
🛡️ Coordinator → it keeps humans in control of the decisions that carry real consequences.
🛡️ Responder → it connects a decision to the person who physically carries it out.
🛡️ Verification → it closes the loop, confirming the outcome rather than assuming it.
That is what "Agents for Humans" means to us. Not an AI that does the most it possibly can, but one that does the right amount - acting when it's clear, asking when it matters, and staying quiet when nothing needs to happen. An agent that turns a rising river into help on the ground, while keeping people informed, understood, and in control the whole way through.
🛡️ Day / Night → it adapts to the real conditions people are in, at any hour.
🛡️ Public Status → it keeps everyone informed, with no login and one source of truth.
🛡️ Coordinator → it keeps humans in control of the decisions that carry real consequences.
🛡️ Responder → it connects a decision to the person who physically carries it out.
🛡️ Verification → it closes the loop, confirming the outcome rather than assuming it.
That is what "Agents for Humans" means to us. Not an AI that does the most it possibly can, but one that does the right amount - acting when it's clear, asking when it matters, and staying quiet when nothing needs to happen. An agent that turns a rising river into help on the ground, while keeping people informed, understood, and in control the whole way through.
👩🏻💻 Working Apps
GitHub - https://github.com/Ajaykumarkv17/ThunAI
YouTube - https://youtu.be/qSHEpTGEAH0
Working App - https://main.d2jea0lxqs5c0.amplifyapp.com/
Series: Agents for Humans: Building thunAI_UGMDU - A Neighborhood AI Agent (3 articles)
- 3Agents for Humans: thunAI_UGMDU in Action - From Early Warning to Real-World Help This article
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article