Night Watch

A two-layer neighborhood safety feed for Kansas City, built because no clean, legal, real-time, geocoded "what's happening near me" feed exists off the shelf (Citizen is a closed ToS-violating scrape; PulsePoint locked its API behind a login in June 2026; KC's open crime data is geocoded but published weeks late).

So Night Watch combines the two halves that are available legitimately:

Layer	Source	Nature
Context	KC Open Data (KCPD), Socrata API	Official, geocoded, but multi-week lag. Weekly digest of crime near home.
Real-time	Broadcastify live scanner feed + local transcription	Live dispatch audio, transcribed on-device with mlx-whisper, matched for near-home streets and priority incidents.

The context layer keeps it honest and grounded; the scanner layer is the actual "alert me now" capability. Neither depends on a service that can revoke access or a ToS we're violating.

The map: crime pins by type, optional live CAD, and live scanner incidents geocoded from alerts, all toggleable from the Layers panel. (Screenshots use a generic downtown location and demo data.)

How it works

crime_pull.py     KC Socrata  ──►  "Neighborhood Watch.md"  (weekly)
scanner_capture.py   Broadcastify ──► spool/*.wav  (rolling 30s segments)
scanner_transcribe.py  spool/ ──► mlx-whisper ──► match ──► mist-notify + "Scanner Alerts.md"

• Capture holds one ffmpeg connection to the feed and writes 16 kHz mono WAV segments. It auto-reconnects on drops.
• Transcribe gates out dead air (RMS energy) so Whisper doesn't hallucinate on silence, transcribes real traffic locally, and matches each line against your near_streets and priority_keywords.
• Matches are spoken via MIST (mist-notify) and appended to a vault note. Rate-limited to one alert per reason per 10 minutes; silent during quiet hours.

Setup

You need two things, both kept out of git:

1. config.json: copy config.example.json to config.json and fill in:

   • home.lat / home.lng: your home coordinates (look up your address on any map, right-click → coordinates). Drives the radius filter. Never committed.
   • radius_miles: how close counts as "near home" (0.75 is a good start).
   • alerts.near_streets: lowercased street names in/around your radius. These are what get matched against spoken dispatch audio. The more complete, the better the geo-filtering.
   • scanner.username / scanner.password: a free Broadcastify account. Needed because free live feeds are account-gated. Or paste a scanner.stream_url_override grabbed from the logged-in web player's Network tab (more reliable; URLs can rotate).

2. Python venv (transcriber only):

   python3 -m venv .venv && .venv/bin/pip install mlx-whisper

   The model downloads on first run and caches.

3. Optional: brew install terminal-notifier so alert notifications are clickable and open the app (otherwise a native notification is used).

Run it

# context layer: try it immediately, no account needed
python3 crime_pull.py

# scanner layer (after config.json has Broadcastify creds):
python3 scanner_capture.py        # terminal 1
.venv/bin/python scanner_transcribe.py   # terminal 2

Desktop GUI

A full-featured control panel (app.py, Flask + pywebview, http://127.0.0.1:5018):

• Dashboard: service health pills (click to start/stop), live transcript stream (feed-tagged), scanner alerts, counters, and a Listen control to tune into any configured feed's live audio in-app (streamed via the backend).

• Map: Leaflet map with your home marker and radius ring. A Layers panel (top-right) toggles each layer on/off: crime pins by type (violent / property / other), the optional Live CAD layer, and the Live Scanner layer: incidents geocoded (approximately) from scanner alerts that drop onto the map as they're heard and refresh every ~10s. Crime pins note their publish lag; scanner pins are marked approximate.
• Settings: geocode your address, set radius, add/remove feeds from the live KC directory, edit near-streets / priority keywords / quiet hours, toggle voice alerts, manage the Broadcastify login. Save and optionally restart capture in one click.

Launch from ~/Desktop/Apps/Night Watch.app, or .venv/bin/python app.py.

Run it as services (launchd)

./install.sh

This fills the launchd/*.plist.template files in for your checkout, writes real plists to ~/Library/LaunchAgents, and loads them. com.nightwatch.crime runs weekly; com.nightwatch.capture and com.nightwatch.transcribe are KeepAlive and run continuously. (The templates use placeholders rather than hardcoded paths, so nothing personal is committed.)

Privacy

Your home coordinates, street list, Broadcastify credentials, captured audio, and transcripts all live in gitignored paths (config.json, spool/, data/, logs/, notify-*.sh). Only config.example.json (generic placeholder coords) is committed. The launchd plists are placeholder templates filled in at install time, so no machine-specific paths are committed either.

Use in another city

Nothing here is hardcoded to Kansas City beyond the example config:

• Crime layer: works with any city's Socrata open-data portal. In config.json set crime.domain and crime.incidents_dataset to that city's crime dataset, and map crime.fields to its column names (the date, offense, address, and location-point columns). Most large US cities publish a Socrata crime dataset; some use ArcGIS instead (a small fetch tweak).
• Scanner layer: Broadcastify covers most of the US (and beyond). Set scanner.directory_url to your metro/county listing page and pick feeds from it in the GUI. Capture and transcription are location-agnostic.
• Alerts: home, radius_miles, and near_streets are all config.

Real-time CAD layer

Where a city publishes a genuinely live calls-for-service feed, Night Watch can pull it as a near-real-time layer (separate from the lagging crime data). Enable it under cad in config.json: set enabled, point domain / incidents_dataset at the live dataset, and map fields to its columns (config.example.json ships a working Seattle example). It polls every 15 minutes (com.nightwatch.cad), shows as the Live CAD map layer, and with cad.alert: true notifies on genuinely new calls near home (deduped by id). KC's own CAD lags, so it's off by default here.

Other data sources worth adding

The same "near home, real-time, on-device" pattern extends to other free, official feeds (good contributions):

• NWS weather alerts (api.weather.gov/alerts): free, no auth, nationwide real-time severe-weather/emergency alerts by point. High value, easy add.
• USGS earthquakes: free real-time GeoJSON feeds.
• NASA FIRMS / Watch Duty: wildfire detection (Watch Duty is West-US).
• State 511 / Open511: traffic incidents and road closures.

Responsible use

This is built for personal, local situational awareness. It captures a free public Broadcastify stream to your own machine and transcribes it on-device. Use it within Broadcastify's Terms of Service: don't rebroadcast or redistribute the audio, and don't republish transcripts. It reads only public open-data and public-safety radio. The transcription is imperfect; never treat an alert as a verified fact or act on it as if it were dispatched to you.

Cost

$0 as configured. Free Broadcastify account, on-device transcription. Broadcastify Premium ($30/yr) would only add ad-free streaming and 365-day archive backfill for overnight gaps. Optional; add it only if the feed proves its worth.

Honest limitations

• Context layer lags weeks. KC publishes crime data late. It's for patterns, not "right now." The note self-labels how far behind it is.
• Scanner geo-filtering is good, not perfect. Dispatchers speak addresses; matching spoken street names from noisy, jargon-heavy audio will miss some and false-positive on others. Tune near_streets and SILENCE_RMS over time.
• Capture needs the Mac awake. Overnight sleep = gaps (Premium archive could backfill, not wired up).