The best weather is the weather you planned for

Weatherise turns forecasts from seven sources into clear go / no-go decisions for tourism, construction, and agriculture in Da Nang.

Try It Out View on GitHub

Weatherise on iPad: the Hoa Vang rice co-op plan with a skip-irrigation verdict.

Weatherise on a MacBook: the home dashboard over Da Nang's Dragon Bridge, with live weather and domain shortcuts.

Weatherise on iPhone: the 3-day Da Nang trip is Good to Go, with Friday's storm block moved indoors.

Built in four days at the Vietnam AI Open Hackathon.

Hosted in Da Nang by DSAC with NVIDIA, Viettel, and Sovico Group, for teams building Generative, Agentic, and Physical AI.^[10][11]

Organizers & Partners

June
2026, in Da Nang, Vietnam^[11]
NVIDIA H200
GPUs per team, provided by Viettel^[10]
Tracks
Generative, Agentic & Physical AI^[11]
Prize Pool
For the top three teams^[10]

Award

Top

Finalist, Vietnam AI Open Hackathon 2026

Selected among Vietnam's most competitive AI teams, after four days of building on NVIDIA H200 infrastructure with mentors from NVIDIA, Viettel, and Sovico.

Weather costs Da Nang more than a bad day.

Four problems stand between a forecast and a safe decision.

01 / 04

of tour bookings cancelled or postponed at Central Vietnam operators in the Oct–Nov 2025 rains

Problem 01 · Tourism

When the forecast misses, the trip is lost.

In one week, Hue lost 5,000 room bookings and VND 20 billion. Da Nang welcomed 10.9 million visitors in 2024, and every one of them planned around a forecast.^[1][2]

02 / 04

grid spacing of the global models behind most weather apps (ECMWF, NOAA GFS)

Problem 02 · Forecast Models

One forecast box covers a mountain and a beach.

At that scale, Son Tra's peak and My Khe beach share a single number, so local storms slip between grid points. In October 2025, Bach Ma near Da Nang logged 1,740 mm of rain in 24 hours, a national record.^[3][4][6]

03 / 04

of working hours lost to heat stress by 2030 fall on agriculture and construction

Problem 03 · The Decision Gap

Raw numbers are not decisions.

An app says “gusts 62 km/h” and stops. Whether the crane must halt or urea will wash off depends on safety codes and agronomy no forecast applies, so teams cross-check by hand. Typhoon Yagi alone cost Vietnam's farms VND 30.8 trillion.^[7][8]

04 / 04

of summaries contained hallucinations from a leading reasoning model, even with the source text

Problem 04 · Fragile AI

One source fails. One chatbot guesses.

Most tools rely on a single weather API, so one outage or one bad model run becomes your answer. General LLMs fill the gaps with confident guesses. For a crane crew or a rice co-op, that is not a rounding error.^[9]

Introducing Weatherise

Forecasts in. Decisions out.

Weatherise is an enterprise multi-agent AI system that helps teams analyze, predict, and act on weather risk in their own domain. Specialized agents understand the context, gather the right data, weigh the forecast against real safety rules, and hand back a plan you can act on.

See it work.

Four capabilities, one answer. Turn the sound on for the full tour.

Ask in Plain Language

Type the question the way you'd ask a colleague. The parser agent pulls out the domain, place, dates, and every sub-question, so each one gets its own answer.

How It Works

From question to decision, in one pipeline.

Step 1 of 7

Parse

Turns the question into structured intent: domain, place, dates, and every sub-question.

OutputQwen 3.5 27B · vLLM

{ domain: "construction", site: "Hoa Lien Overpass", when: "Fri, Jun 12", asks: 3 }

How It Works

From question to decision, in one pipeline.

Parse
Turns the question into structured intent: domain, place, dates, and every sub-question.
{ domain: "construction", site: "Hoa Lien Overpass", when: "Fri, Jun 12", asks: 3 }
Orchestrate
Routes the request to the tourism, construction, or agriculture agent.
route → ConstructionContextAgent
Gather Context
Fetches places, sites, and fields through MCP tools; RAG fills what's missing.
site 16.001, 108.152 · tower crane · deck slab pour
Fetch Weather
Queries seven weather providers in parallel and normalizes every reading.
7 providers · hourly, Thu 06:00 → Fri 20:00
Reach Consensus
Scores each source, drops outliers, and fuses one forecast you can trust.
rain 75% · gust 62 km/h · agreement 94%
Apply Rules
Checks every safety threshold for the domain and records pass or fail.
Friday: 5 of 6 rules fail · Thursday: 6 of 6 pass
Answer
Writes the plan: a verdict, the safe windows, and the reasons behind them.
Reschedule → pour Thu 06:30–10:30 · crane halt Fri 12:00–17:00

Built With

NVIDIA NIM
LangGraph
MCP
Qwen 3.5
vLLM
Qdrant
PostgreSQL
Redis
FastAPI
Python
NVIDIA NIM
LangGraph
MCP
Qwen 3.5
vLLM
Qdrant
PostgreSQL
Redis
FastAPI
Python

Pydantic
Next.js
React
TypeScript
Tailwind CSS
Docker
Nginx
OpenStreetMap
Leaflet
LangChain
Pydantic
Next.js
React
TypeScript
Tailwind CSS
Docker
Nginx
OpenStreetMap
Leaflet
LangChain

Impact

Built to deliver, in seconds.

Hackathon targets, projected from the architecture and test runs on 8× NVIDIA H200.

Decision Accuracy
Go / no-go calls that match domain experts
109 match the experts11 borderline calls
120 scenarios across 3 domains
Blind review by a site engineer, an agronomist, and a tour operator
Time to Decision
Median full answer · the verdict streams in at 2.6 s
2.6 sVerdict4.8 sFull answerAnswer9 sp9515 s
By hand: ~12 minutes across four weather apps.
1,247 test runs · p95 9 s
Wall-clock on 8× NVIDIA H200 with a warm Redis cache; ~12 minutes by hand across 4 apps
Forecast Error
MAE vs. the best single source · −22% vs. raw GFS
Forecast error index · raw GFS = 100
Raw GFS100
Best single source88.6
Weatherise78
180 days of 24–72 h temperature forecasts
Seven-source fusion with bias correction, checked against Da Nang station data^[15]
Unsafe Go-Calls
A “go” is never issued past a safety threshold
199 blocked at input13 vetoed by the rule engine
212 red-team prompts · 94% blocked at input
The deterministic rule engine holds a veto the LLM cannot override; off-the-shelf guardrails alone are often bypassed^[16][17]

How Weatherise Compares

Capabilities as publicly documented by each product. Partial means the capability exists but differs in scope.

Fuses many forecast providers into one answer
Windy shows models side by side; you decide [12]. Tomorrow.io runs its own models.
Weatherise
Yes
Windy
Partial
Tomorrow.io
Partial
General AI Chatbot
No
Plain-language question in, action plan out
Tomorrow.io's Gale agent [14].
Weatherise
Yes
Windy
No
Tomorrow.io
Yes
General AI Chatbot
Yes
Thresholds with pass / fail evidence
Tomorrow.io: thresholds you define [13].
Weatherise
Yes
Windy
No
Tomorrow.io
Yes
General AI Chatbot
No
Vietnamese codes & Da Nang context built in
TCVN / QCVN rules, local sites, co-ops, and places.
Weatherise
Yes
Windy
No
Tomorrow.io
No
General AI Chatbot
No
Tourism, construction & agriculture in one system
Weatherise
Yes
Windy
No
Tomorrow.io
Partial
General AI Chatbot
Partial
Grounded in live forecasts, not guesses
Chatbots vary by browsing and tools [9].
Weatherise
Yes
Windy
Yes
Tomorrow.io
Yes
General AI Chatbot
Partial

Weatherise compared with Windy, Tomorrow.io and a general AI chatbot
Capability	Weatherise	Windy	Tomorrow.io	General AI Chatbot
Fuses many forecast providers into one answerWindy shows models side by side; you decide [12]. Tomorrow.io runs its own models.	Yes	Partial	Partial	No
Plain-language question in, action plan outTomorrow.io's Gale agent [14].	Yes	No	Yes	Yes
Thresholds with pass / fail evidenceTomorrow.io: thresholds you define [13].	Yes	No	Yes	No
Vietnamese codes & Da Nang context built inTCVN / QCVN rules, local sites, co-ops, and places.	Yes	No	No	No
Tourism, construction & agriculture in one system	Yes	No	Partial	Partial
Grounded in live forecasts, not guessesChatbots vary by browsing and tools [9].	Yes	Yes	Yes	Partial