Machines that know when not to trust themselves
TerraGuard watches a real ecosystem the way an AI safety layer watches a model. Five Earth-sphere models, fused through a central Trust Core into one honest stress readout, live across Korea.
Being confidently wrong is dangerous.
A model that is sure of itself and wrong can knock out a grid, mislead a student, or mismanage an ecosystem. The Trust Core answers that in three moves.
Calibrate
Every output ships with a split-conformal interval, a statistically valid range with a coverage guarantee behind it.
Explain
Feature attribution on every prediction, so you can see which inputs drove the score, in plain language.
Abstain
When inputs drift or the ensemble disagrees, the model withholds rather than assert a number it cannot vouch for.
The same layer that governs a language model now governs an ecosystem.
I am an interdisciplinary builder obsessed with trustworthy AI, machines that know when not to trust themselves.
And I use it to watch over a world I love, starting with the DMZ.
The five spheres
Every past project becomes a data stream
Sun
F10.7 solar-flux forecasting with split conformal intervals. Space weather that can knock out a grid.
Atmosphere
Active-fire and air-quality hazard over the corridor, the fastest-moving acute risk we watch.
Hydrosphere & Biosphere
Watershed chemistry from a field logger and canopy health from multispectral UAV flights.
Geosphere
Net carbon flux and the CaCO3 sequestration work that models how much can be pulled back.
See it running
Open the live monitor for the whole peninsula, or step into the dashboard where the five spheres fuse into one governed readout.