Environmental risk, built into products that global enterprises run.

What it does

• Three fitted models, thirteen engineered layers. Gradient-boosted trees for flood susceptibility. A frequency-ratio model for landslide susceptibility. A transparent least-squares fit for urban heat. The other layers are calculated indices, official warnings republished unchanged, or reference geometry.
• Two audiences, one grid. An underwriter gets district and 1 km exposure, and can attach a return period to it. A resident gets a named street in their own language, with the reason attached: low ground, far from a working drain.
• Sixteen hazards across two regions. India covers flood, heat, dams and rivers, groundwater, drought, landslide, forest loss, cloudburst and air quality. The Middle East covers dust storms, wet-bulb heat, wadi flash floods, cyclones and air quality, across nine Gulf cities.
• Every layer states its own method on the map: fitted model, index, rule, or an official warning republished unchanged. An underwriter always knows what they are reading.

How well it works

The first two scores are ROC-AUC, where 0.5 is a coin toss and 1.0 is perfect. Landslide reads 0.85 on Himachal Pradesh and 0.81 on Uttarakhand. Each state is scored by the model that did not train on it. Heat is R², the share of observed variance in land surface temperature that the model explains.

What it does

• Two mechanisms, two models. The fluvial model learns from Sentinel-1 radar change over several real floods per city. The pluvial model learns from documented waterlogging hotspots, which the satellite can barely see.
• Every score carries its reasons. Minto Bridge in Delhi comes back as a pronounced sink, far from a drain. That is exactly why the underpass floods.
• Past monsoons replay on the map, with the recorded rainfall against each flood event. A route check also flags known flood points along a journey.
• A new city is a config file, not new code — one YAML plus a hotspot list.

How well it works

Both columns hold out spatial blocks inside each city. That is a stricter test than a random split. The pluvial column uses positive–unlabeled learning, scored against sampled background. It measures how well the model retrieves known hotspots, not calibrated probability. The two columns are not comparable to each other.

What it does

• Nine series, seven viruses. RSV, influenza split into A and B, rhinovirus, hMPV, seasonal coronavirus, adenovirus and parainfluenza. Only influenza and RSV have an established forecasting literature. The rest are largely unmodelled.
• A standardised 0 to 100 score for each virus, region and horizon. It forecasts case load. That is why it is called risk, not demand.
• Two horizons, two purposes. One to four weeks is the operational range, and where the measured skill sits. Six months is a planning view. It ships as a probability band, not a point number.
• Quantile forecasts, not point estimates. The ensemble produces a full forecast distribution with conformal intervals. The intervals are scored alongside the central path.

How well it works

The scoring rule is WIS, the Weighted Interval Score. The United States CDC uses it to rank forecasts in its FluSight challenge. Each series is scored against the strongest naive baseline for that series, pooled over ten regions, with expanding-window walk-forward testing. WIS grades the whole forecast distribution. It therefore measures where the intervals sit and how wide they are, not only the central path. The values below are the percentage improvement on that baseline.

Eight of the nine series beat the baseline. Parainfluenza does not, and it stays on the table. Its season is weak and irregular next to the other viruses, which makes a persistence baseline hard to beat. Seven of these series have no published benchmark at all. That is why the comparison is against a naive baseline.

What it does

• Four sensor families feed it. Text comes from conflict-event databases, thermal from NASA FIRMS, radar from Sentinel-1 SAR, and optical from Sentinel-2. The radar and optical detectors run locally, not as a vendor call.
• Entity resolution across space and time. One strike can appear as a news item, a database entry and a satellite change, at slightly different places and times. The engine groups candidates by place, time and type, then scores each pair. It merges too readily on purpose and then splits back, because a missed event costs more than a duplicate.
• Confidence is independence-weighted. A noisy-OR combines the source families, and a correlation matrix discounts sources that move together. Ten outlets carrying one wire story therefore count as one source. A radar detection over the same cell counts as a real second.
• Ranked, not listed. About 11,900 events are sorted by severity, corroboration and recency. A fourth term measures how unusual each event is against the normal rate for that grid cell.

Validation

• The radar damage detector was checked against United Nations satellite damage assessments of Mariupol.
• Burn scars and open water are the classic confusion pair in optical change detection, because both are dark in the near-infrared. Over the 2022 Sviati Hory fires the detector returned 33 burn detections. It misclassified none of them as water.
• The pipeline is deterministic: an automated test rebuilds the whole map from the raw records and checks it is identical, bit for bit.

Built not to be misused

• Locations round to about 1 km before anything is stored. Exact coordinates are discarded at the door. The tool can say "near Avdiivka". It cannot produce a point to aim at.
• Decision support, never targeting. The rule is enforced in code where data leaves the system. The data model also cannot hold a person at all.