Frontier Analytics

Navigation

FrontierAnalyticsDisaster intelligence · Philippines

For the night the river comes up.

Stations that read the water as it rises. Risk scores for every address in its path. One live map for the people running the response. Scroll through a single night in Naga City — first rain to first light.

Watch the nightTalk to the engineers

~20 typhoons enter Philippine waters a year · PAGASA

22:35Before the rain

Start with the ground.

Naga sits on the Bicol plain, a couple of metres above the rivers that flood it. Before a drop falls we have already joined the elevation model, the national hazard layers and every building footprint — into one map you can query by address.

GLO-90 elevation · PHIVOLCS · UP NOAH · footprints

23:50The city, asleep

The hazard science exists. It arrives too late.

PHIVOLCS, UP NOAH and the JRC publish excellent hazard science — at provincial resolution, as a PDF, after the event. We build the layer that makes it usable while it still matters.

  • At the riverbankin a datacenter
  • Per addressper province
  • During the stormin the report after

01:20Landfall · the Naga River bursts its banks

01Sense

Stormwatch

It keeps reading when the grid goes down.

River gauges are scarce, and the ones that exist go quiet with the power. Stormwatch is a solar camera that finds the floodwater, reads its depth and grades its severity on the device — day, night, and through the storm.

  • Raspberry Pi 5
  • Hailo AI accelerator
  • YOLOv11 segmentation
Inside Stormwatch

02:45The water at its worst · one address, read out

02Assess

Project Bedrock

The flood map, down to one front door.

Bedrock scores the address itself — what floods it, what shakes it, what it is like to live there, and an expected loss with its uncertainty attached. One lookup, or a whole loan book overnight.

  • JRC depth-damage curves
  • PHIVOLCS
  • UP NOAH
Inside Project Bedrock

04:10The response · city hall to the worst-hit streets

03Respond

Haribon

One live map, instead of forty group chats.

Haribon puts incidents, sensor readings, hazard scores, stock and vehicles on one live map — so the people dispatching are reading a situation instead of reassembling it from messages.

  • GIS
  • MongoDB
  • AWS
Inside Haribon

05:48First light

Sense. Assess. Respond.

Each product stands on its own. Together they cover the whole night — from the first reading at the riverbank to the last truck home at dawn.

  1. 01Sense

    Stormwatch

    It keeps reading when the grid goes down.

  2. 02Assess

    Project Bedrock

    The flood map, down to one front door.

  3. 03Respond

    Haribon

    One live map, instead of forty group chats.

The working

We show our working.

One reading from one station, carried line by line until someone can act on it. Every line has a figure, and every figure can be checked.

Worked example · illustrative values

  1. Figure 1: side elevation of a Stormwatch station — solar panel, printed enclosure with a Raspberry Pi 5 and Hailo accelerator, and two cameras looking down at the river.Stormwatch station · side elevationNot to scalePi 5 · HailoSolar panelno grid power neededPETG chassisprinted in-houseInference on boardPi 5 + Hailo acceleratorLocal queuebackfills on reconnectTwo camerasone NoIR, for the darkPowerSolarComputeOn deviceServiceHand tools
    Fig. 1 · What stands on the riverbank

    Fig. 1Engineering

    The grid is down. On the bypass bridge, SW-03 takes a frame at

    01:24:07

    Built to outlast the outage. The network is the first thing a typhoon takes out, so the station computes its own answer: inference on a Hailo accelerator, power from a panel, readings queued until the link comes back. The chassis is printed in-house and opens with hand tools.

    Raspberry Pi 5 · dual cameras, one NoIR

  2. Figure 2: a camera frame of a flooded street, with the water segmented and its depth read off a doorway as 1.42 metres.SW-03 · NoIR camera01:24:07 PHT1.42 mWater in frame51 %Depth1.42 m ± 0.11SeverityAlert
    Fig. 2 · Floodwater mask and depth from a single frame

    Fig. 2Research

    51 % of that frame is floodwater, so the depth is

    1.42 m ± 0.11

    The depth reading is peer-reviewed. YOLOv11 segmentation finds the water in the frame; surface-normal estimation recovers the geometry of the scene, which turns extent into depth. No staff gauge in shot, no one standing in the water.

    Co-authored with Melchor Filippe S. Bulanon

  3. Figure 3: the JRC depth-damage curve for residential buildings in Asia. At 1.42 metres of water the damage ratio is reported as a band, 0.55 to 0.65.Depth → damage · residential, AsiaJRC · Huizinga et al. 20170123456 m0.00.51.01.42 m → 0.55–0.65a band, not a false decimalHazardkept separateLivabilitykept separateLosswith its band
    Fig. 3 · From water depth to expected damage

    Fig. 3Methodology

    At 1.42 m, a house on this street takes damage of

    0.55–0.65

    Every score has a curve behind it. Depth becomes damage through the JRC depth–damage curves, over PHIVOLCS, UP NOAH, HazardHunterPH and GLO-90 elevation. We report a band, not a decimal the elevation model can’t support — and hazard and livability stay separate numbers.

    Disclosure-ready for BSP Circular 1085

  4. Figure 4: a live response map. The river has flooded two roads, and the route from the command post to the evacuation centre has been redrawn around them.Haribon · live map · Naga04:10 · illustrativeNaga RiverCommand postEvacuation centreRoads cut2RouteredrawnScreens in syncevery one
    Fig. 4 · One map for everyone running the response

    Fig. 4Operations

    By 04:10, on every screen in the command post at once:

    2 roads cut

    The map redraws itself. Who is affected, which roads are cut, how to reach the evacuation centre — recomputed as readings and reports land, and the route is redrawn around the water. Nobody should be hitting refresh in the middle of a disaster.

    MongoDB · AWS · Convex, reactive end to end

2 people, end to end

Small team. No hand-offs.

The same people solder the stations, train the models, write the scoring method and answer your email. Nothing between the riverbank and the loan book is outsourced — so when something is still weak, we know exactly where, and we say so.

One continuous line from the river to an envelope, passing through a flood station, a laptop training a model, a flood staff beside a house, and an envelope: soldering the station, training the model, writing the method and answering your email are done by the same two people.Solder the stationTrain the modelWrite the methodAnswer your email

One line, riverbank to inbox. Never lifted.

Meet the team

Signal No. 3 over your worst night

Bring us your worst night.

Mail goes straight to the engineers who built this. Tell us what you have to decide, and by when. Every blank you fill, the storm lets up a little.

To the engineers · team@frontieranalytics.ph

I for . By , I need to know .

Or start with Stormwatch