wx.jamestannahill.com

Live hyperlocal weather dashboard for Midtown Manhattan, New York. Data from a private Ambient Weather WS-2902 station, updated every 5 minutes.

Live: wx.jamestannahill.com · API: api.wx.jamestannahill.com/current

---

Features

Dashboard

• Live conditions: temperature, feels-like, humidity, dew point, wind, gusts, pressure, UV index, solar radiation, rainfall
• Anomaly headline: how today compares to the station's own rolling baseline for this month and hour
• Percentile rank: where current temperature falls in the station's historical distribution
• Yesterday's summary: prose recap of the prior day (high/low, rain events, notable winds), auto-generated nightly
• Today in History: daily verdict panel backed by 156 years of NOAA Central Park records and 85 years of ERA5 reanalysis. Shows whether today's high and low are hotter/cooler/normal vs. the historical average, with deviation bars spanning p5→p95 in actual °F and a tick at the long-run mean
• WeatherKit forecast: Apple WeatherKit 7-day outlook with daily high/low, condition, precipitation probability, and UV index
• NWS alerts: active National Weather Service warnings for the station coordinates (via api.weather.gov)
• Nearby stations: Weather Underground personal weather stations in the area (the home station excluded) with current conditions for local comparison
• Comfort calendar: 30-day heatmap grid, color-coded by daily comfort score (0 to 100)
• Station records: current-month extremes: temp high/low, max gust, peak rain rate, pressure range, each with the date it was set
• Analog forecast: +1h/+2h/+3h predictions via nearest-neighbor pattern matching on 90 days of station data, with running MAE accuracy
• Urban Heat Island delta: this station vs. the average of JFK, LGA, and EWR METAR readings, updated every 5 minutes
• Rain probability: logistic regression model fitted weekly on labeled station history, with persistence boost if currently raining

Chart

• Multi-field selector: Temperature, Humidity, Wind, Pressure, Urban Heat
• Time ranges: 12h / 24h / 7d / 30d with automatic downsampling
• Historical baseline overlay (dotted line)
• ±1σ anomaly bands: shaded region between ±1 standard deviation of the baseline, derived from Welford's online variance computed across 90 days of readings
• Drag-to-zoom, wheel/trackpad zoom centered on cursor, two-finger pinch on mobile, double-click to reset
• Copy the rendered chart to the clipboard as a branded PNG (title, location, timestamp baked in)
• Share uses the Web Share API on mobile (image + URL) and falls back to clipboard + Twitter intent on desktop
• Rendered with uPlot for high-performance canvas drawing
• Shipped as a Preact island (client:visible): ~90 KB of chart JS only loads when the chart scrolls into view

API

Public, read-only, no authentication required.

Endpoint	Description
GET /current	Latest reading with all ML signals and metadata
GET /history?hours=N	Last N hours of readings (max 720), downsampled for longer ranges, with per-slot baselines and std dev
GET /rain-events?days=N	Parsed rain events from the last N days (default 30)
GET /daily-summaries?days=N	Pre-computed daily summaries from the last N days (default 30)
GET /nearby	Latest snapshot of nearby personal weather stations
GET /health	Health check: data freshness and poller status
GET /forecast	Standalone analog forecast
GET /records	Standalone station records
GET /climate-context	NOAA/ERA5 climate context
GET /alerts	Active NWS alerts for station point
GET /export?days=N&format=csv	Bulk export (json or csv)
GET /openapi.json	OpenAPI 3 specification

The analog forecast, station records, and daily summary are embedded in the GET /current response (forecast, station_records, daily_summary); there are no standalone endpoints for them.

---

Architecture

Ambient Weather WS-2902
        │  (Wi-Fi, every 5 min)
        ▼
Ambient Weather Cloud API
        │
        ▼
wx-poller (Lambda, EventBridge 5-min)
   • Fetches reading
   • Range + stuck-sensor validation
   • NOAA METAR fetch → UHI delta
   • Welford rolling stats (mean + variance per MM-HH slot)
   • Writes to wx-readings (90-day TTL)
        │
        ├──► wx-forecaster (Lambda, 30-min)
        │      • Nearest-neighbor analog forecast on 90-day hourly data
        │      • Evaluates previous forecast accuracy → MAE stored
        │
        ├──► wx-ml-fitter (Lambda, weekly Sun 3AM UTC)
        │      • Scans 90-day readings, labels rain events
        │      • Fits logistic regression w/ class weighting
        │      • Stores coefficients to wx-ml-models if F1 ≥ 0.05
        │
        ├──► wx-summarizer (Lambda, daily 5AM UTC)
        │      • Processes last 30 days (idempotent)
        │      • Computes daily stats + prose summary
        │      • Writes to wx-daily-summaries
        │
        ├──► wx-records-tracker (Lambda, weekly Sun 2AM UTC)
        │      • Scans 90-day readings
        │      • Computes per-month station records
        │      • Writes to wx-station-records
        │
        ├──► wx-alerter (Lambda, 15-min)
        │      • Evaluates anomaly thresholds
        │      • Sends SES email on trigger (2h cooldown)
        │
        ├──► wx-climate-bootstrap (Lambda, one-time)
        │      • Backfills 156yr NOAA Central Park daily records (noaa-ghcn-pds S3)
        │      • Backfills 85yr ERA5 reanalysis via Open-Meteo
        │      • Writes to wx-climate-normals
        │
        └──► wx-climate-updater (Lambda, daily 6AM UTC)
               • Appends yesterday's NOAA + ERA5 observations
               • Keeps climate baseline current

wx-api (Lambda, API Gateway HTTP API → CloudFront)
   • /current  : live reading + all ML signals + climate context + daily summary + records
   • /history  : downsampled readings with baselines + std dev
   • /rain-events: parsed rain events
   • /daily-summaries: pre-computed day summaries
   • /nearby   : latest nearby-stations snapshot

Cloudflare Worker (Astro 6 SSR) → wx.jamestannahill.com  (dashboard)
   • Hero + 12 conditions cards SSR'd at request time from /current
     → first paint shows real values, AI crawlers see the whole dashboard
   • Edge-cached 60s (s-maxage=60, stale-while-revalidate=300)
   • /og.png generated on demand at the edge via Satori + resvg-wasm
     (workers-og); a real snapshot keyed by ?v= is edge-cached 5 min,
     a degraded render (API miss) is served uncached so the next
     crawler scrape recovers the current snapshot; no AWS dependency
   • Chart shipped as Preact island, hydrates on client:visible
   • Partytown moves GA off the main thread; /docs prefetched on hover

API Gateway + CloudFront → api.wx.jamestannahill.com

---

Tech Stack

Layer	Technology
Station	Ambient Weather WS-2902, Wi-Fi gateway
Compute	AWS Lambda (Python 3.12, arm64): API + ingest only
Scheduling	AWS EventBridge
Storage	AWS DynamoDB (on-demand, 12 tables)
API	AWS API Gateway HTTP API
CDN	AWS CloudFront (API only); Cloudflare Workers (dashboard + OG)
Dashboard	Astro 6 server-rendered on Cloudflare Workers, Preact island for the chart, vanilla JS for everything else, uPlot, NHG Display font; design rules in astro/DESIGN.md
OG image	Generated at the edge via workers-og (Satori + resvg-wasm), bundled NHG Display TTF
IaC	AWS CDK (Python)
Email	AWS SES
Secrets	AWS Secrets Manager
Climate data	156yr NOAA GHCN-Daily (Central Park) + 85yr ERA5 reanalysis
Forecast	Apple WeatherKit API
Active alerts	NOAA National Weather Service API
Airport METAR	NOAA aviationweather.gov

---

DynamoDB Tables

Table	Key	Purpose
wx-readings	station_id + timestamp	Raw 5-min readings, 90-day TTL
wx-daily-stats	station_id + month_hour	Rolling mean + variance per MM-HH baseline slot
wx-alerts	alert_type	Alert debounce state
wx-forecasts	station_id	Latest pre-computed analog forecast
wx-forecast-accuracy	station_id + evaluated_at	Per-evaluation MAE + running mean
wx-uhi-seasonal	station_id + month	Monthly rolling UHI delta averages
wx-ml-models	model_id	Fitted logistic regression weights + metrics
wx-daily-summaries	station_id + date	Daily stats + prose summary
wx-station-records	station_id + month	Per-month extreme records with dates
wx-climate-doy	mmdd	NOAA GHCN per-day-of-year distributions (156yr Central Park)
wx-climate-hourly	doy + hour	ERA5 per-DOY-hour distributions (85yr reanalysis)

---

ML Details

Rain Probability

Logistic regression: p = σ(w·x + b). Features: humidity (normalized), 1-hour pressure trend, dew-point depression, sin/cos of hour-of-day. Weekly refit via gradient descent (300 epochs, L2 regularization, class weighting n_neg/n_pos). Current model: F1=0.396, recall=82%, n=24,348.

Analog Forecast

1-hour buckets of the last 90 days → normalize to [0,1] → 6-hour fingerprint → nearest 5 Euclidean neighbors → average their next 3-hour trajectories. Confidence = inverse of mean neighbor distance, normalized to 0 to 100%.

Baseline Variance (±1σ bands)

Welford's online algorithm accumulates per-slot variance alongside the rolling mean. Transition to EMA variance at 8,640 samples (30 days). Standard deviation is stored as std_{field} per MM-HH slot and served via /history.

Today in History (Climate Context)

On each /current call, climate_context.py loads the NOAA + ERA5 normals for today's calendar date and computes a daily_verdict for both temp_high and temp_low. Each verdict includes: the observed value, the long-run p5/p50/p95 percentiles in °F, a z-score, and a label (much hotter / warmer / normal / cooler / much colder). The dashboard renders deviation bars that start at the historical average and extend to today's value, with the full p5→max(p95, value) range as the track.

---

Repo Structure

lambdas/
  shared/           # DynamoDB client, Secrets Manager, UHI fetch
  wx_poller/        # 5-min station data collection
  wx_api/           # REST API + all signal computation
  wx_alerter/       # Anomaly alert emails
  wx_bootstrap/     # One-time ERA5 + WU backfill (readings)
  wx_climate_bootstrap/ # One-time NOAA GHCN + ERA5 climate normals backfill
  wx_climate_updater/   # Daily NOAA + ERA5 climate update
  wx_forecaster/    # Analog forecast + accuracy tracking
  wx_ml_fitter/     # Weekly logistic regression training
  wx_summarizer/    # Daily prose summaries
  wx_records_tracker/ # Weekly station records
astro/
  src/pages/
    index.astro     # SSR dashboard (hero + conditions grid pre-rendered)
    docs.astro      # Prerendered (export const prerender = true)
    og.png.ts       # Edge-rendered OG image (workers-og + Satori)
    sitemap.xml.ts  # Dynamic sitemap (lastmod = render date)
    404.astro       # Branded 404 page
  src/middleware.ts # Security headers (HSTS, nosniff, frame, referrer) on all SSR responses
  src/layouts/      # Base.astro: shared head, schema, fonts
  src/components/
    Chart.tsx       # Preact island for uPlot chart (client:visible)
    SchemaOrgIndex.astro / SchemaOrgDocs.astro
  src/og-fonts/     # NHG Display TTFs bundled into the worker for OG render
  public/           # app.js, style.css, llms.txt, _headers, weatherkit.png
  astro.config.mjs  # cloudflare adapter, partytown, preact, prefetch
cdk/
  wx_stack.py       # CDK stack (Lambdas, DynamoDB, API Gateway, API CloudFront)
  app.py            # CDK entry point
scripts/
  deploy_worker.sh  # Build Astro, patch wrangler.json, wrangler deploy

---

Deployment

# Deploy AWS infrastructure (Lambdas, DynamoDB, API Gateway, etc.)
cd cdk && npx cdk deploy --require-approval never

# Deploy dashboard Worker to Cloudflare
#   - Builds Astro (SSR for /, prerender for /docs, served from docs.html)
#   - Patches the auto-generated wrangler.json (name, vars, custom domain)
#   - wrangler deploy → wx.jamestannahill.com
bash scripts/deploy_worker.sh

First-time setup needs npx wrangler login (one-shot OAuth) and a DNS record for wx.jamestannahill.com that doesn't conflict with the Custom Domain.

Estimated cost: about $4 to $6 per month AWS, Worker on the free tier.