FrontierAnalytics · An AI and data company · Manila, Philippines
We build the models, sensors and data infrastructure the Philippines uses to get ahead of climate disaster — and to price the property standing in its way. Preparedness and response for the teams on the ground. Hazard and property intelligence for the institutions carrying the risk.
PAGASA · ~20 tropical cyclones enter the PAR each year
Scroll
What we do
Climate preparedness and response for agencies, LGUs and the teams that move first. Property and hazard intelligence for banks, insurers and developers. Both run on the same sensing, the same national hazard layers, the same address-level models.
Climate response · Property intelligence
PHIVOLCS, UP NOAH and the JRC publish world-class hazard science. It arrives as a PDF, at provincial resolution, after the fact. Our work is the layer on top of it — computer vision on the riverbank, address-level risk models over the national hazard data, one live operational picture — so the number exists while there is still time to use it.
Edge AI
models that run at the station, not in a datacenter
Address
resolution at the house, not the province it sits in
Live
an answer during the event, not a report after
Nobody knew how high it got.
River gauges are scarce, and the ones that exist go down with the grid. Stormwatch is a solar camera that measures water level and severity on the device — day, night, and through the storm.
Raspberry Pi 5 · Hailo AI accelerator · YOLOv11 segmentation
A province-wide map, pricing a single house.
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
The response is being run in a group chat.
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
The loop · Sense → Assess → Respond
Stormwatch senses: computer vision on the flood line, running at the station. Bedrock assesses: address-level hazard and expected loss for the property in its path. Haribon responds: one live map the operation actually runs on. Each one is useful alone; together they cover the event end to end, which is the only way the number ever reaches the person who has to act on it.
How we work
We are an AI company that shows its working. A segmentation method that went through peer review. A scoring rule written down, including where it stops being precise. Hardware we print and service ourselves. Check all of it before you trust anything it outputs.
01 · Research
The method is published
02 · Methodology
A score that survives an audit
03 · Engineering
Designed for the outage
04 · Operations
One map, every screen
Research.
Before the station existed there was a paper. YOLOv11 segmentation paired with surface-normal estimation turns ordinary camera footage into a flood depth — no staff gauge, no technician in the water. It went through peer review, so you can check the working.
Co-authored with Melchor Filippe S. Bulanon
Methodology.
JRC depth-damage curves over PHIVOLCS, UP NOAH, HazardHunterPH, GLO-90 elevation and NSCP seismic zones. Hazard and livability stay separate. Precision is capped, because a fourth decimal place the elevation model cannot support is a lie with a number in front of it.
BSP Circular 1085 · disclosure-ready
Engineering.
The network is the first thing a typhoon takes out, so the station computes its own answer: inference on a Hailo chip, power from a panel, a PETG chassis printed in-house and serviceable with hand tools.
Raspberry Pi 5 · dual cameras, one NoIR
Operations.
A real-time GIS platform on MongoDB, AWS and Convex. Who is affected, what is still reachable, how to get there — updating on every connected screen at once, because nobody should be hitting refresh during a disaster.
Batch pipeline + API
A small engineering team in the Philippines. We solder the stations, train the models, and write the scoring method ourselves — nothing between the flood line and the loan book is outsourced, which is also why we can tell you exactly where it is still weak.
Edge AI · geospatial data · risk modelling — built here
A loan book to score before disclosure season. A river that needs watching. A response plan for the next landfall. Mail goes straight to the engineers who built this, and we answer it ourselves.