Skip to content
vnmoorthyPublic

About

Ondo 温度: the AI that makes heat safe for every body. Cruise guests at a Penang onsen just talk, in any language; a deterministic safety engine plus live, cited You.com data gives each a personal Soak Prescription, and staff watch everyone in the water. Built on Emergent, voice by Willow.

Topics

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

4 Commits

Folders and files

Repository files navigation

Ondo 温度: the AI that makes heat safe for every body

Ondo 温度

The AI that makes heat safe for every body.

Ondo is the AI heat-safety host for Yu Pinang Onsen in Penang. Cruise passengers just talk, in any language, about who they are, how they feel and the medicines they take. Ondo checks Penang's live heat with You.com, runs a deterministic safety engine, and gives every guest a personal Soak Prescription, a guided soak and a staff team that is watching.

Built with Emergent Grounded by You.com Voice by Willow Speech to Ship 2026
FastAPI React MongoDB License: MIT Safety guidance, not medical advice

▶ Live demo  ·  Pitch deck  ·  3-minute storyboard  ·  10-minute demo script  ·  Architecture  ·  Safety engine  ·  How we chose the idea

Soak safely in Penang's heat.


Contents


The 20-second version

Margaret, 71, from Sydney, steps off Navigator of the Seas into a 34°C Penang afternoon. She holds fn and says:

"I'm Margaret, 71, from Sydney. I take metoprolol for my blood pressure and I put a seasickness patch on this morning. I'd love a long hot soak."

Ondo's risk dial swings to red. "Your seasickness patch (scopolamine) stops you sweating. In a 41°C bath on a 34°C Penang day, your body can't cool down." Ondo moves her from the 41°C Hinoki bath to the 37.5°C Jade bath for 6 minutes, with cited sources from You.com. Her name lights up on the Staff Safety Console with a personal timer. Then a heat spike hits, and every guest's limit tightens live.

We just prevented a heat stroke, in her language.


Why Ondo exists

Heat is a silent, growing killer, and a hot bath stacks heat on heat.

The fact Source
🌡️ About 546,000 heat-related deaths a year in the last decade (average for 2012–21). Lancet Countdown 2025, via PIK Potsdam · report
👵 Heat deaths among people over 65 were 167% higher in 2023 than in the 1990s. Lancet Countdown 2024
♨️ Japan estimates ~19,000 bathing-related deaths a year, mostly people over 65, from "heat shock". Official advice: water under 41°C, soaks of at most 10 minutes. Consumer Affairs Agency, via nippon.com

Cruise passengers are the highest-risk onsen guests.

  • They are often older, and on blood-pressure pills, beta-blockers or diuretics.
  • Many wear scopolamine seasickness patches, which stop sweating.
  • They are dehydrated after days at sea and the ship's drinks package.
  • They step from a 22°C ship into 34°C, humid Penang, and then into a 41°C bath.

The onsen knows none of this about its guests. It puts up a sign. Its staff can't read 20 languages. Ondo is the host that knows.


What it does

Feature What the guest or staff member gets
🎙️ Just talk Hold fn and speak with Willow, tap the mic, or type, in any language. The LLM turns rambling speech into editable profile chips: age, medicines with their generic names and drug classes, conditions, today's factors (drinks, sleep, seasickness, sunburn), pregnancy, children, language and ship.
🌤️ Live heat check You.com Search with knowledge: "core" returns George Town's live temperature, humidity and heat index, shown with a Live · You.com badge (or Cached when it falls back).
🧭 Soak Prescription A deterministic Python rules engine sets the bath, max °C, minutes per round, rounds, cool-downs and water. A 0–100 risk dial runs jade → amber → vermilion.
🚩 Cited red flags Every red-flag card has a plain-language reason and evidence chips from the You.com Answer API, with links to the sources.
🎫 Soak Pass Booking into real slots, a pass code like OND-7K2Q, a QR code, the ship line ("all aboard 20:30 · leave the onsen by 18:45").
⏱️ Guided soak Per-round countdowns and cool-downs with spoken check-ins in the guest's language ("Margaret, two minutes left. Sit up slowly before you stand."), plus a hydration tracker.
🆘 "I feel unwell" One big button alerts staff instantly and shows calm steps in the guest's language: get out slowly, sit, sip water, staff are coming.
🩺 Health summary A generated summary for the ship's doctor, in English plus the guest's language.
🖥️ Staff Safety Console Every guest in the water with time in water versus personal limit, risk colours, red flags, next check-in, an alerts rail with acknowledge, live conditions, today's ships and impact KPIs. Auto-refresh every 3 s.
🗣️ "Say what's happening" Staff hands are wet, so staff talk: "Margaret looks dizzy" or "Rotenburo is 42 degrees right now." The LLM turns it into an incident: flag the guest and show the protocol, or lower the bath's temperature and re-check everyone in it.
🔥 Heat-spike simulation Heat index 46°C re-runs the engine for every guest. Limits tighten, alerts fire and announcements are spoken.
💬 Ask Ondo "Can I soak after two beers?" Grounded answers from You.com with citations, translated into the guest's language.
📱 Judges join live A QR on the staff console opens the guest app; a new guest appears on the board within 3 seconds.

Screenshots

Captured from the live app at https://heat-safe-onsen.emergent.host (Ask Ondo panel is an illustration).

Just talk Soak Prescription Soak Pass
Just talk intake Soak Prescription Soak Pass
Staff Safety Console Body heat curve Ask Ondo
Staff Safety Console Body heat curve Ask Ondo

How it works

Ondo system architecture

flowchart LR
    subgraph People["People + voice"]
        G["Cruise guest<br/>React guest app"]
        S["Onsen staff<br/>Staff Safety Console"]
        W["Willow dictation<br/>+ Web Speech fallback"]
    end

    subgraph API["FastAPI backend · /api"]
        I["Intake parser<br/>LLM → Pydantic profile"]
        C["Conditions service<br/>live heat index"]
        R{{"Safety rules engine<br/>deterministic · unit-tested"}}
        E["Evidence service<br/>cited answers"]
        X["Explainer + translator<br/>LLM, never changes numbers"]
        B["Bookings + Soak Pass"]
        A["Soak sessions + alerts"]
        N["Staff incident parser"]
        D["Demo controls<br/>heat spike · 10× · reset"]
    end

    subgraph Ext["Grounding, models + data"]
        YS[("You.com Search<br/>knowledge: core")]
        YA[("You.com Answer<br/>citations")]
        L[("Emergent Universal LLM key<br/>Claude Sonnet 5.5")]
        M[("MongoDB")]
    end

    W -. types into .-> G
    W -. types into .-> S
    G -->|"talk"| I --> R
    C --> R
    R --> E --> X --> G
    G --> B --> A
    A -->|"alerts"| S
    S -->|"say what's happening"| N --> A
    D --> R
    C --- YS
    E --- YA
    I --- L
    X --- L
    N --- L
    B --- M
    A --- M
    R --- M
Loading

The loop, end to end: talk → profile chips → live heat → rules engine → prescription with cited red flags → book → Soak Pass → guided soak with spoken check-ins → "I feel unwell" → the staff console shows the guest and the alert → a staff voice incident → heat-spike re-evaluation for everyone.

Full sequence diagrams, caching and fallbacks are in docs/ARCHITECTURE.md.


The safety engine: rules decide, the LLM explains

The numbers that keep a guest safe are never produced by a language model. They come from a small, readable, unit-tested Python rules table (app/backend/rules.py). The LLM's jobs are to understand messy speech, to explain the result warmly in the guest's language, and to parse staff incidents. It cannot change a limit.

Baseline: max 41°C · 10 min per round · 3 rounds · 5 min cool-down · 250 ml water before and after each round · risk 10. Every matching rule applies its limit; the most restrictive limit wins and the risk points add up.

Trigger Max °C Max min Rounds Notes Risk +
Age 65–74 40 8 3 enter slowly, no cold plunge right after 15
Age 75+ 39 6 2 staff check each round 25
Scopolamine patch / anticholinergics 38 6 2 indoor or shaded only, +500 ml 30
Beta-blockers 40 8 – stand up slowly, no cold plunge 15
Diuretics – 8 – +500 ml 15
Vasodilators, nitrates, PDE5, ACE/ARB 39 8 – no alcohol 15
Anticoagulants – – – handrails, slip care 5
Insulin or diabetes 39 8 – glucose check, eat first 15
Pregnancy 37.5 10 2 private bath · first trimester: red 25
Heart disease 37 5 1 clearance · heart attack < 6 months: red 40
Epilepsy – – – never alone, private bath with companion 20
Alcohol 1–2 drinks (last 3 h) 39 6 – wait 1 hour if possible 15
Alcohol 3+ drinks red: no hot bath, foot bath only 40
Poor sleep / seasick today – 8 – +500 ml 10
Children under 7 38 5 – adult supervision 15
Live heat index ≥ 40°C −1°C on every cap −2 min – shaded or indoor 10
Live heat index ≥ 45°C indoor gentle baths only 20

Risk level: below 35 is jade (green), 35–60 is amber, above 60 or any red rule is vermilion (red). Red means the gentle bath or foot bath only, and staff confirm before the guest soaks. Ondo recommends the bath whose temperature is closest to, but not above, the guest's max °C, honouring privacy needs (private family bath for pregnancy, epilepsy, tattoos or modesty).

The six baths of Yu Pinang Onsen (seed data): Hinoki indoor 41°C · Rotenburo outdoor rock bath 40°C (shaded) · Jade gentle bath 37.5°C · Cool plunge 18°C · Private family bath 37–40°C adjustable (tattoo-friendly, modest) · Step-free accessible bath 38°C.

Full table, scoring, bath selection, evidence queries, worked examples and limitations: docs/SAFETY_ENGINE.md.


You.com integration

You.com is live and load-bearing: without it there is no live heat index and no cited evidence. Every call is server-side only, sends X-API-Key and an explicit User-Agent: Ondo/1.0, is cached in MongoDB for 10 minutes, and falls back to seeded data with a Cached badge so the demo never breaks.

Feature Endpoint Query
Live heat index for George Town Search POST https://ydc-index.io/v1/search with knowledge: "core" George Town Penang weather now heat index humidity
Evidence for each red flag Answer POST https://api.you.com/v1/answer How does {medication or condition} affect heat tolerance and hot bath or sauna safety?
Guest questions ("Can I soak after two beers?") Answer the guest's question + (hot spring onsen safety)
Ships in Penang today (staff console) Answer Which cruise ships are in Penang Swettenham Pier on {date}, with arrival and departure times?
Heat news (staff console) Search, freshness week Penang heatwave heat warning MetMalaysia

How the results are used. Evidence text comes from contents.highlights[0], then snippets[0], then description. Temperature, humidity and "feels like" are parsed from the knowledge description with a regex, with an LLM fallback. Answer's [[1, 2]] markers index into citations and render as small source chips that link out. The common red-flag queries are pre-warmed at startup, because Answer takes about 3 s.

Timeouts: 12 s for Search, 20 s for Answer. On a timeout or any non-200 response the API returns "live": false and the UI shows Cached instead of an error. Seeded fallback conditions: 33°C, 75% RH, heat index 41°C.


Built by voice with Willow

Voice is the interface, for the people who use Ondo and for the people who built it.

  • Older guests can't type drug names on a phone, but they can say them, in 100+ languages. The intake box is one big text area labelled "Just talk: who you are, how you feel today, and any medicines. Hold fn and speak with Willow, or tap the mic." Willow types straight into it; a browser Web Speech mic is the fallback, with a visible listening state.
  • Staff hands are wet. The console's "Say what's happening" box is built for Willow dictation, so a staff member can report an incident without touching a keyboard.
  • The build itself was spoken. The product spec and the prompts to Emergent's agent were dictated with Willow. Willow works at the OS level, so it types into Emergent's chat, the deployed app, or anything else with a text field.

Tips for demoing with Willow: add scopolamine, metoprolol, warfarin, Rotenburo, Hinoki, Swettenham, Yu Pinang and Ondo to your Willow personal dictionary, and keep the text visible before you submit so you can fix a misheard drug name. The chips step is the read-back: guests confirm their medicines before the plan is built.


Built on Emergent

The whole two-sided, full-stack app was built by Emergent's agent from our spoken prompts and spec:

  • React guest app and Staff Safety Console, mobile-first, large type and big tap targets for older guests.
  • FastAPI backend with a unit-tested rules engine, Pydantic-validated LLM output, server-side You.com calls and caching.
  • MongoDB for guests, prescriptions, bookings, sessions, alerts and incidents, so everything survives a refresh.
  • Emergent Universal LLM key with Claude Sonnet 5.5 for extraction, explanation, translation and incident parsing.
  • Public emergent.host deploy: https://heat-safe-onsen.emergent.host.

The app source lives in app/ (app/frontend and app/backend), exported from Emergent.


How we chose the idea

Team V9 drew the brief at Speech to Ship (no re-rolls): "A hot springs onsen for cruise ship passengers with one day in port in Penang, Malaysia."

We generated 30 ideas across six angles and ran a tournament. A simulated panel of about 70 agents, modelled on the public roles of this event's judges (Emergent, Willow, You.com, Bond AI) plus VC, design and voice-AI archetypes, scored each idea on the official rubric. Twelve simulated cruise-passenger and operator personas reacted to each idea, and a red team attacked build risk, local fit and venture risk. These are simulations, not endorsements by real people.

Finals (top 10, average out of 5):

# Idea Finals avg Vision One line
1 LastCall 4.39 3.08 An airport departure board for people in bathrobes.
2 SoakClock 4.15 4.00 Soak until the last safe minute, guaranteed back on board.
3 SoakRx → Ondo 3.89 3.08 Speak for 30 seconds and get a soak prescription that knows your meds, today's Penang heat and your all-aboard time.
4 YuTurn 3.83 3.58 Tell it your all-aboard time in any language, soak, and make the U-turn back to your ship.
5 ShipDrop 3.72 5.00 Every ship gets its own onsen drop: passengers pool before they land.
5 Rempah-Yu 3.72 2.00 Speak a memory; it becomes a spice bath today and a jar you carry home.
7 Bibik-san 3.60 2.00 Half Nyonya matriarch, half ryokan okami: a voice host who teaches you to bathe right.
8 PierRunner 3.46 3.42 If you're going to miss the ship, we already know how you'll catch it.
9 Berthcast 3.39 4.50 Every ship that berths becomes a demand forecast, a price ladder and a staffing plan.
10 SafeSoak 3.31 2.75 A cited, minute-by-minute soak protocol for today's Penang heat.

Why SoakRx became Ondo. The simulated judges called its spoken scopolamine-patch catch a "gasp moment" and "real empathy". Their critique was just as useful: medication advice reads as liability unless the venue owns it, and they wanted the onsen and staff side "front and centre" with rules deciding and the LLM only explaining. Ondo is that critique, built: a venue-side Staff Safety Console, a deterministic rules engine, cited evidence and a read-back of every medicine. Full method, all 30 ideas and scores: docs/TOURNAMENT.md.


Business model

Who pays What they buy Why
Venues: onsen, spas, saunas, thermal baths Per-guest SaaS, about US$0.50–1 per guest, or a monthly plan It replaces a laminated sign and closes a liability gap.
Insurers Lower premiums for venues that run Ondo; Ondo shares the saving Fewer incidents and an audit trail.
Cruise lines Licensing for ship spas, medical centres and shore excursions They already know their guests' risk profiles.
Consumers and employers (next) A personal heat-safety subscription for people over 65 and their families; heat-safety compliance for employers of outdoor workers The same engine, pointed at everyday heat.

The trillion-dollar path

Wedge → engine → category.

  1. Wedge. One Penang onsen shows that cruise guests soak more safely, stay longer and spend more.
  2. Thermal venues worldwide. Japan's ~27,000 onsen facilities, saunas, spas and thermal baths, plus cruise-ship spas.
  3. Personal heat safety. The same engine tells anyone what their body can take today: older adults, outdoor workers, athletes, travellers. Billions now live through dangerous heat every year.
  4. Category. The health layer of climate adaptation, where the ~US$6T+ wellness economy, health insurance and climate resilience meet.

Make heat safe for every body.


Run it locally

Ondo is a standard Emergent stack: FastAPI + MongoDB + React. The app source lands in app/ from Emergent; the exact commands below will be matched to app/ once the code lands (check app/backend/requirements.txt and app/frontend/package.json).

Prerequisites: Python 3.11+, Node 18+ with Yarn, MongoDB 6+ (local or Atlas), a You.com API key and an Emergent Universal LLM key.

git clone https://github.com/vnmoorthy/ondo.git
cd ondo
cp .env.example .env              # then fill in the values below
Variable Used by What it is
MONGO_URL backend MongoDB connection string, e.g. mongodb://localhost:27017
DB_NAME backend Database name, e.g. ondo
YDC_API_KEY backend You.com API key (sent as X-API-Key, never to the browser)
EMERGENT_LLM_KEY backend Emergent Universal LLM key (Claude Sonnet 5.5)
REACT_APP_BACKEND_URL frontend Where the React app finds the API, e.g. http://localhost:8001

Backend

cd app/backend
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
uvicorn server:app --reload --host 0.0.0.0 --port 8001
pytest -q                          # runs the rules-engine unit tests

Frontend

cd app/frontend
yarn install
yarn start                         # http://localhost:3000

Open http://localhost:3000 for the guest app and http://localhost:3000/staff for the Staff Safety Console. With no YDC_API_KEY, Ondo runs on its seeded fallbacks and shows Cached badges.


API reference

All routes are under /api. The first five are defined in the product spec; the rest describe the planned surface and will be confirmed against app/backend once the code lands.

Method Path Purpose
GET /api/conditions Live George Town temperature, humidity and heat index from You.com Search (live: true/false)
POST /api/intake {text} → structured guest profile (LLM, Pydantic-validated)
POST /api/prescription {profile} → rules-engine prescription, risk score and level, red flags with You.com evidence, explanation in the guest's language
POST /api/bookings Book a bath and slot → pass code like OND-7K2Q
POST /api/ask Grounded answer with citations from You.com Answer, translated
GET /api/baths, /api/slots The six baths and 30-minute slots from 13:00 to 19:00 with capacity
GET /api/pass/{code} Pass, prescription summary and ship line
POST /api/pass/{code}/… Start soak, log water, "I feel unwell"
GET /api/pass/{code}/health-summary Summary for the ship's doctor (English plus the guest's language)
GET /api/staff/board Guests in the water, time versus limit, risk, red flags, next check-in
GET / POST /api/staff/alerts, /api/staff/alerts/{id}/ack Alerts rail and acknowledge
POST /api/staff/incident {text} → incident (flag a guest, or lower a bath's °C and re-check everyone in it)
GET /api/staff/kpis, /api/ships Impact KPIs; ships in Penang today
POST /api/demo/heat-spike, /api/demo/fast-forward, /api/demo/reset Demo drawer

Example:

curl -s -X POST "$REACT_APP_BACKEND_URL/api/intake" \
  -H 'Content-Type: application/json' \
  -d '{"text":"I am Margaret, 71, from Sydney. I take metoprolol and I put a seasickness patch on this morning."}'

Data model

MongoDB collections (names will be matched to app/backend once the code lands):

Collection Holds
guests Profile from intake: name, age, language, home country, ship, party, medications (as said, generic name, class), conditions, today's factors, pregnancy weeks, children's ages
prescriptions Engine output: bath, max °C, minutes, rounds, cool-down, water, risk score and level, matched rule ids with reasons and evidence, explanation
baths The six baths with temperature, indoor/outdoor, privacy and accessibility flags
slots / bookings 30-minute slots with capacity; bookings with pass code, guest, bath and time
sessions Guided soaks: round, time in water, personal limit, hydration log, next check-in
alerts Limit reached, "I feel unwell", heat-index spikes; acknowledged by and when
incidents Staff voice reports and the structured action taken
conditions / api_cache You.com responses cached for 10 minutes, keyed by endpoint and normalized query

Safety & privacy

Ondo gives safety guidance, not medical advice. It does not diagnose, treat or replace a doctor. It applies conservative, published-guidance-style limits to help an onsen keep guests safe, and staff check on every guest. When in doubt, Ondo says "don't soak today" and asks staff to confirm.

  • Rules decide, the LLM explains. Limits come from a deterministic, unit-tested table. The LLM cannot raise a limit; its output is validated with Pydantic.
  • Every red flag is cited. Evidence comes from the You.com Answer API with visible sources.
  • Read-back. Guests see and edit what Ondo heard (medicines especially) before a prescription is built.
  • Health data stays minimal. Ondo asks only what it needs for today's soak and never asks for an ID or a card. API keys stay on the server. The guest's words go to the LLM (through Emergent's Universal LLM key) to build the profile and the explanation; You.com evidence queries carry a medicine or condition name, never the guest's name.
  • Production checklist for venues. Get explicit consent at intake, set a short retention window, restrict the staff console to staff, and follow local health-data law (in Malaysia, the PDPA).
  • Voice. Willow is the speaker's own dictation app and types text into Ondo; the browser mic fallback uses the Web Speech API. Ondo receives and stores text, not audio.

Roadmap

  • Real screenshots and a 60-second demo video in this README
  • Consent screen and per-venue data-retention setting
  • Medication dictionary expansion with pharmacist review, and a "we didn't recognise this medicine" path that always asks staff
  • Staff sign-in and an exportable incident audit trail for insurers
  • Wearable and water-temperature sensor input for the console
  • Multi-venue dashboard for spa groups and cruise-ship spas
  • Personal heat-safety mode for everyday heat: older adults, outdoor workers, athletes

Contributing

Ideas, bug reports and pull requests are welcome. Please read CONTRIBUTING.md first; changes to the safety rules need a source and a unit test.

License

MIT © 2026 Team V9 (Ondo).

Acknowledgements

  • Emergent for hosting Speech to Ship at Emergent HQ, San Francisco (Nov 20, 2026), and for the agent that built this app.
  • You.com for the Search and Answer APIs that ground every claim.
  • Willow for making voice the keyboard.
  • Bond AI for co-running the event.
  • The Lancet Countdown, PIK Potsdam, Japan's Consumer Affairs Agency and nippon.com for the data that made the problem impossible to ignore.

Ondo 温度 · Yu Pinang Onsen · Penang
Make heat safe for every body.

About

Ondo 温度: the AI that makes heat safe for every body. Cruise guests at a Penang onsen just talk, in any language; a deterministic safety engine plus live, cited You.com data gives each a personal Soak Prescription, and staff watch everyone in the water. Built on Emergent, voice by Willow.

Topics

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages