Skip to content

FrederikSchorr/aiwindy

BranchesTags

Repository files navigation

AIWindy — AI-Powered Sailing Weather Advisor

Live App: aiwindy.schorr.wien

An AI-powered sailing weather analysis app for European waters. Each analysis presents 5 interactive Windy maps paired with AI-generated explanations — zooming in from the continental scale down to the local sailing area:

  1. Druck & Luftmassen — Europe-wide 850hPa pressure and air mass map (ECMWF)
  2. Fronten — KNMI weather fronts analysis chart with AI interpretation
  3. Wind & Welle — Regional high-resolution wind model map (1–7 km) with forecast bullets
  4. Wolken & Regen — Cloud and precipitation overlay with thunderstorm risk
  5. Temperatur — Local meteogram with temperature summary

The app combines data from European synoptic services (Meteonews, Wetterzentrale, KNMI), national weather APIs (GeoSphere Austria, DHMZ Croatia, EMY/OpenSkiron Greece), and multiple AI models to generate concise, sailor-relevant analysis. Users can also upload photos or videos for meteorological AI analysis.

Architecture Overview

Client (React + Vite)
  └── POST /api/chat (SSE)
        ├── Message classification (GPT-4.1-mini)
        ├── Location detection (Claude Sonnet)
        ├── European weather data
        │   ├── Meteonews (general overview text)
        │   ├── Wetterzentrale (850 hPa charts with local timestamps)
        │   └── KNMI (front analysis charts with local timestamps)
        ├── National weather data (per country)
        │   ├── AT: GeoSphere Austria JSON APIs
        │   ├── HR: DHMZ XML feeds
        │   └── GR: EMY gale warnings + OpenSkiron WRF 4km GRIB
        ├── LLM preprocessing (Claude Haiku/Sonnet)
        └── Weather output generation (5 sections, Claude/GPT)
  └── POST /api/upload (SSE)
        ├── Photo: EXIF GPS/timestamp → GPT-4.1 Vision analysis
        └── Video: ffprobe metadata + ffmpeg thumbnail → Gemini 2.5 Flash analysis

The backend pipeline is JSON-based: raw data → preprocessed structured text → AI-generated output sections. Each stage writes to a persistent analysis JSON file (analyses/).

Message Classification

GPT-4.1-mini classifies each user message:

  • ANALYSE <location> — Location-specific weather query → full 5-section analysis
  • CHAT — General meteorology/sailing question or follow-up → direct GPT-4.1 answer (with full last analysis context: meta, sections, preprocessed data)
  • UNCLEAR — Ambiguous query → asks user to specify location

Location Detection & Windy Model Selection

Claude Sonnet detects sailingArea + city from user input, matched against data/sailingareas.json (133 sailing areas across 20 countries). The full sailing area list is sent as system prompt with Anthropic prompt caching (cache_control: ephemeral) to reduce cost and latency. Geocoding via Nominatim.

Windy wind model is selected via static JSON lookup — no LLM call:

  1. data/sailingareas.json — each sailing area has windyModel (highest priority)
  2. data/countries.json — country-level fallback
  3. server/location.ts — coordinate-based fallback for unlisted countries

Models: aromeHd (1.3km), czeAladin (2.3km), ukv (2km), iconEu (7km), gfs (22km).

European Weather Data

Source URL Data File
Meteonews meteonews.at/de/Allgemeine_Lage/K33/Europa Allgemeine Wetterlage Europa (HTML scrape) server/weather-europe.ts
Wetterzentrale wetterzentrale.de/maps 850hPa ECMWF Temperatur-/Druckkarten (Bild-URLs) server/weather-europe.ts
KNMI cdn.knmi.nl/.../weerkaarten Frontenanalyse + Frontenprognose (GIF charts) server/weather-europe.ts
Windy embed.windy.com Interaktive Karten: 850hPa, Wind, Wolken, Meteogramm Frontend iframes

National Weather Data

Country Source URL Data File
🇦🇹 Austria GeoSphere Austria dataset.api.hub.geosphere.at Hourly wind, gusts, temperature, precipitation, cloud cover (JSON API) server/weather-national-austria.ts
🇦🇹 Austria Austrocontrol austrocontrol.at/wetter/wetter_fuer_alle/wettervorhersage Flugwetter Wetterlage + FXOS Vorhersage (HTML scrape) server/weather-national-austria.ts
🇦🇹 Austria LSZ Burgenland lsz-b.at/fuer-buergerinnen/sturmwarnung-webcams/ Neusiedler See Sturmwarnungen (HTML scrape) server/weather-national-austria.ts
🇭🇷 Croatia DHMZ prognoza.hr/jadran_h.xml Adria Segelwetter-Vorhersage (XML API) server/weather-national-croatia.ts
🇭🇷 Croatia DHMZ prognoza.hr/pomorci.xml Maritime Warnungen, Wind, Seegang, Sicht (XML API) server/weather-national-croatia.ts
🇭🇷 Croatia DHMZ prognoza.hr/sedam/hrvatska/7d_meteogrami.xml Städte-Meteogramme, Temperatur (XML API) server/weather-national-croatia.ts
🇬🇷 Greece HNMS/EMY newportal.hnms.gr/emy/.../naftilia_deltio_thalasson_ektiposi Sturmwarnungen, Seewetter-Bulletin (HTML scrape) server/weather-national-greece.ts
🇬🇷 Greece OpenSkiron openskiron.org/en/openwrf WRF 4km GRIB: Wind, Welle (Douglas), Wolken, Temperatur, CAPE server/weather-national-greece.ts + openskiron_fetch.py

Other countries (20 total in sailingareas.json): analysis uses Europe-wide data + Windy maps. LLM preprocessing pipeline available for future integrations.

OpenSkiron WRF 4km (Greece)

Python subprocess (scripts/openskiron_fetch.py) handles GRIB1 fetch and extraction since GRIB parsing requires native ecCodes.

python scripts/openskiron_fetch.py <domain> <wind_lat> <wind_lon> <city_lat> <city_lon> [grib_url]
  1. Node.js discovers current GRIB URL, checks DB cache (URL-based invalidation), passes URL as optional 7th arg
  2. Downloads + bz2-decompresses if not cached in cache/openskiron/; .url sidecar detects staleness
  3. Opens GRIB1 with cfgrib (non-standard WRF table → uses parameter IDs, not shortNames)
  4. Extracts 49-step hourly time series at nearest grid point → JSON to stdout; results cached in DB per coordinate pair

4 domains cover all Greek sailing areas (openskiron_domain in sailingareas.json): Ionian_Islands_4km, Aegean_SW_4km, Aegean_NW_4km, Aegean_SE_4km

Weather Data Preprocessing

Two-stage LLM preprocessing (server/weather-national.ts):

  1. National synopsis (preprocessNationalWeather) — extracts/translates the country-level weather overview (e.g. DHMZ Adria synopsis, EMY marine bulletin, Austrocontrol Wetterlage)
  2. Local weather (preprocessLocalWeather) — extracts area-specific data: warnings, wind, wave, cloud/rain/thunderstorm, temperature. Each country module has dedicated functions (e.g. extractDhmzWarning, preprocessGreeceLocalWind)

LLM calls (Claude Sonnet / GPT-4.1-mini) extract and translate from raw XML/HTML into structured German text. Numeric data (GeoSphere JSON, OpenSkiron GRIB) is preprocessed without LLM.

Weather Output Generation

Single Claude Sonnet 4.6 call (server/weather-output.ts) with all preprocessed data + weather chart images (Wetterzentrale 850hPa, KNMI fronts). Returns a JSON object with 5 section texts (bullet-point style). Each section has strict rules for content scope, bullet count, and word limits.

Caching

Layer Key Pattern Storage Invalidation File
Location lookups loc:{normalized input} PostgreSQL cache_store No TTL (permanent) server/cache-db.ts
OpenSkiron weather data openskiron:sa:{domain}:{coords}, openskiron:city:{domain}:{coords} PostgreSQL cache_store URL-based (new GRIB URL → cache miss) server/cache-db.ts
OpenSkiron GRIB files cache/openskiron/{domain}.grb2 Filesystem .url sidecar file detects staleness scripts/openskiron_fetch.py
Analysis results analyses table (JSONB) PostgreSQL Append-only (1 row per analysis) server/analysis-store.ts

Photo & Video Upload

Camera button in chat input. Accepts images (JPEG, PNG, WebP, HEIC) and videos (MP4, QuickTime, WebM), max 20MB.

Feature Photos Videos
Metadata EXIF extraction (exif-parser) → GPS + timestamp ffprobe → GPS (ISO6709) + creation_time from MP4 atoms
AI Analysis GPT-4.1 Vision Gemini 2.5 Flash
Thumbnail Original image ffmpeg: 1s frame extraction, "▶ Video" overlay
SSE event { exifMeta } { videoMeta: { thumbnailBase64, time, locationName, countryCode } }

Analysis sections: 📷 Aufnahme, ☁️ Wolkentyp, 🌧️ Regen, 🌊 Wellen, 🌫️ Bedeckungsgrad, 🌤️ Wetterentwicklung. Videos additionally: 💨 Windgeschwindigkeit.

If GPS found: reverse geocodes via Nominatim, shows location + date, offers "ja" button to trigger weather analysis for that location.

Key Files

server/
  routes.ts                    Express API endpoints (chat, upload, geocode, KNMI proxy)
  weather-europe.ts            European data (Meteonews, Wetterzentrale, KNMI) + time helpers
  weather-national.ts          National dispatch (AT/HR/GR) + preprocessing pipeline
  weather-national-austria.ts  GeoSphere Austria + Austrocontrol + LSZ Burgenland
  weather-national-croatia.ts  DHMZ Croatia integration
  weather-national-greece.ts   EMY + OpenSkiron Greece integration
  weather-output.ts            AI output generation (single Claude Sonnet call → 5 sections)
  analysis-store.ts            Analysis JSON persistence (filesystem + PostgreSQL)
  location.ts                  Location detection (sailing area + city via Claude Sonnet)
  cache-db.ts                  PostgreSQL cache helpers (TTL, location cache, OpenSkiron cache)

data/
  sailingareas.json            133 sailing areas with windyModel, coordinates, openskiron_domain, emy_name
  countries.json               Country-level wind model fallback
  windymodels.json             Windy model definitions (key + label)
  windsystems.json             Local wind systems per country (Bora, Meltemi, etc.)

scripts/
  openskiron_fetch.py          Python: GRIB1 fetch + extraction for Greece

client/src/pages/
  home.tsx                     Single-column chat UI with progressive section rendering

Frontend

Single-column chat interface built with React + Tailwind + shadcn/ui. Progressive rendering via SSE:

SSE event UI update
{ location } Header + Section 1 (Windy 850hPa map, no marker)
{ weatherEurope } Sections 2–5 (KNMI fronts chart + 3 Windy iframes)
{ weatherOutput } Bullet text fills in for all 5 sections

Chat mode (CHAT classification) uses GPT-4.1 with full last analysis context (meta, sections, preprocessed data) for follow-up questions.

Setup

Requirements

  • Node.js 20+ with npm
  • Python 3.10+ with pip (for OpenSkiron/Greece only)

Installation

npm install
pip install -r requirements.txt

Environment Variables

DATABASE_URL=postgresql://...
SESSION_SECRET=...
AI_INTEGRATIONS_ANTHROPIC_API_KEY=sk-ant-...
AI_INTEGRATIONS_OPENAI_API_KEY=sk-...
AI_INTEGRATIONS_GEMINI_API_KEY=...

Development

npm run dev

Custom Domain

  • Domain: aiwindy.schorr.wien
  • Redirect from .replit.app/.replit.dev to custom domain in production

About

Windy weather charts for sailors, explained with open data by AI

aiwindy.schorr.wien/

Topics

weather sailing europe

Resources

Readme
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages