Conflict Map

A fully local, news-driven world map of current armed conflicts: who is fighting, who is backing whom, and what the consequences are.

How it works

1. GDELT – every 15 minutes GDELT publishes a global event file. We keep the "material conflict" events (CAMEO quad class 4) for the last 48 h and use them as a background heat layer (per-country fill + heatmap) and for actor-pair links.

2. RSS feeds – ~50 feeds (see app/config.py): BBC/Guardian/France 24 regional desks, NYT, WaPo, CNN, NPR, CBC, Sky, DW, Euronews, Politico EU, VOA, UN News, ReliefWeb, Google News keyword feeds (which surface Reuters and AP), regional outlets (Kyiv Independent, Ukrainska Pravda, TASS, Times of Israel, JPost, Haaretz, Middle East Eye, Al-Monitor, Anadolu, Africanews, AllAfrica, Dawn, The Hindu, SCMP, Japan Times, Yonhap, The Diplomat) and defence/analysis sites (Defense One, Breaking Defense, The War Zone, War on the Rocks, Long War Journal, Crisis Group, Bellingcat). Feeds are fetched in parallel. Articles whose title and summary contain none of the RELEVANCE_TERMS are stored but never sent to the model.

3. Local LLM extraction – unprocessed articles go to gpt-oss-120b on llama-swap (http://127.0.0.1:8080/v1) in batches of 40, together with the existing conflict records. The model returns updated/new conflict records: parties (side + role), consequences by category, latest developments with source citations, an epicenter.

4. Map – MapLibre GL 5 (vendored) in globe projection with atmosphere, or flat Mercator via the Globe toggle; idle globe slowly spins. Basemap is fully local: dark-tinted shaded relief with seabed, built from Natural Earth GRAY_50M_SR_OB by scripts/build_tiles.py into static/tiles/ (512px tiles, z0-4, ~4 MB), plus 50m country polygons, lakes, rivers and admin-1 lines, and Open Sans glyphs in static/fonts/ for country/capital/city labels. Conflict markers sized by severity and coloured by status; dashed arcs from supporters/mediators to the epicenter; click a conflict (or a country) to focus: parties keep their side colour, everything else darkens.

5. Attacks – the model also lists every located attack reported in each batch (weapon incl. shelling/ground/bombing, attacker, optional origin, target, launched/intercepted counts, outcome, source), for internal conflicts as well as cross-border ones. Attacks without a usable origin pulse at the target instead of drawing a trajectory.

6. Incidents – GDELT fight / mass-violence events with coordinates, aggregated per place over the window, drawn as dots from zoom 3 with the most-mentioned source article on hover/click. In focus mode only incidents inside the selected conflict's countries are shown. This is press attention, not a verified incident log. Place names are resolved through a local gazetteer (static/data/places.geojson, Natural Earth 10m populated places, ~7,300 cities); the model's own coordinates are only accepted when they fall inside the named country, otherwise the strike snaps to the country centroid with precision country. Precision is city, approx or country and is shown in the UI. On the map each strike replays as an animated projectile along a bent trajectory with an impact flash; a day slider filters to one day.

7. Military aircraft – the server polls the public ADS-B aggregator airplanes.live (/v2/mil, falling back to adsb.fi) once a minute and keeps the history in memory. /api/aircraft only ever serves a snapshot at least AIRCRAFT_DELAY_MIN (20) minutes old, with the last AIRCRAFT_TRAIL_MIN (30) minutes as a trail, so the map never shows live positions near fighting. Aircraft are coloured by role from their ICAO type code (tanker, surveillance, transport, combat, helicopter). Only aircraft that broadcast ADS-B appear, which in practice means support and surveillance flights, not combat missions. This runs on the public viewer too (it does not go through the database); after a restart the layer stays empty until the first delayed snapshot is available.

Everything lives in data/conflict.db (SQLite). Nothing leaves the machine except the fetches to GDELT, the RSS feeds and the ADS-B aggregator.

Rebuild the relief tiles (only needed if you change the colours in the script):

curl -LO https://naciscdn.org/naturalearth/50m/raster/GRAY_50M_SR_OB.zip && unzip GRAY_50M_SR_OB.zip -d /tmp/gray
uv run --with pillow --with numpy python scripts/build_tiles.py /tmp/gray/GRAY_50M_SR_OB.tif

Run

As a service on the home machine (restarts on crash and at boot):

cp deploy/conflict-map.service ~/.config/systemd/user/   # once
systemctl --user enable --now conflict-map
loginctl enable-linger $USER                            # keep running when logged out
journalctl --user -u conflict-map; tail -f data/server.log

By hand:

uv sync
uv run uvicorn app.server:app --host 127.0.0.1 --port 8765
# open http://127.0.0.1:8765

The server refreshes every 30 min (REFRESH_MINUTES). Trigger a refresh manually with the Refresh button or curl -X POST localhost:8765/api/refresh.

Run the pipeline by hand:

uv run python -m app.pipeline            # GDELT + feeds + LLM extraction
uv run python -m app.pipeline --skip-llm # just fetch data

Env vars: LLM_BASE, LLM_MODEL, GDELT_BACKFILL_HOURS (48), GDELT_WINDOW_HOURS, REFRESH_MINUTES (30), ARTICLES_PER_BATCH (40), AIRCRAFT_ENABLED (1), AIRCRAFT_DELAY_MIN (20), AIRCRAFT_TRAIL_MIN (30), AIRCRAFT_POLL_SECONDS (60).

Conflict stats

The top of each conflict panel shows how long it has lasted, attacks reported this week, press-reported clashes and outlets reporting. Below that, "Figures in recent reports" lists death tolls, displacement etc. as verbatim quotes from the linked articles, with who reported them, the period covered and a "party claim" tag when a warring side is the source. app/stats.py reads the full text of recent articles; a figure is kept only if its number and quote appear in the article. Refreshed per conflict at most every 12 h.

Public viewer (Luma001)

The Hetzner box has no GPU, so it only serves the map. The home machine runs the pipeline and pushes a WAL-safe snapshot of data/conflict.db after every refresh (PUSH_TARGET=luma:/root/conflict-map/data, set in run.sh; the file is swapped in atomically with mv). The container runs with SERVE_ONLY=1: read-only database, no pipeline loop, /api/refresh returns 403, the Refresh button is hidden.

/root/conflict-map          git clone of this repo (public), docker compose, port 172.17.0.1:8030
/root/caddy/Caddyfile       conflicts.lumatechsolutions.co.uk { reverse_proxy 172.17.0.1:8030 }

Deploy an update: ssh luma 'cd /root/conflict-map && git pull --ff-only && docker compose up -d --build'.

Notes / known limits

• GDELT is noisy and English-media-biased (the US always glows). The heat layer only uses CAMEO root codes 19/20 (fight, mass violence) and a power curve, but it is still a rough "attention" signal, not a casualty count.
• The LLM decides what counts as a conflict; the prompt excludes crime, politics and diplomatic disputes but it will occasionally over-reach. Every development cites the article it came from so claims can be checked.
• Conflicts are never deleted automatically. last_seen is stored; stale ones can be filtered in the UI later.

Military ships layer

Naval vessels from AIS via aisstream.io, limited to watched seas (SHIPS_BOXES in app/config.py) and held back 20 minutes like aircraft. It needs a free API key from https://aisstream.io; without one the layer shows "(off)". On the server put it in /root/conflict-map/.env as AISSTREAM_API_KEY=... and run docker compose up -d. Most warships switch AIS off; auxiliaries (USNS, RFA), patrol and coast-guard vessels are what usually shows up.

Contributing

Issues and pull requests are welcome: https://github.com/Comm4nd0/conflict-map

The public site is a read-only viewer; the data pipeline (feeds, GDELT, local LLM extraction) runs on a separate machine. To work on it locally, run ./run.sh with a local OpenAI-compatible model endpoint (LLM_BASE, LLM_MODEL), or run the viewer alone against any data/conflict.db with SERVE_ONLY=1.

License

Code: MIT, see LICENSE. Third-party data keeps its own terms: Natural Earth (public domain), GDELT, news articles (linked and quoted, not redistributed), ADS-B via adsb.fi / airplanes.live, AIS via aisstream.io. MapLibre GL JS (BSD-3-Clause) and Open Sans (OFL) are vendored under static/.