Detect changing relationships
Watch how current data and learned assumptions diverge—not just whether a single score moves.
Research / Adaptive intelligence
Learning when the world changes.
My doctoral research studied how a model can continuously adapt as the relationship between historical and real-time data evolves.
- Online Meta-Learning
- Model Adaptation
- Deep Learning
- Smart Grid
Research questionWhat should a model do when
What should a model do when
yesterday stops predicting today?
Static training assumes the future will resemble the data a model already knows. Real systems rarely stay that tidy.
My research explored continuous model adaptation using online meta-learning: a framework for learning how to update a deep model as real-time data arrives and underlying patterns change.
The smart-grid application made that question concrete. Loads, behaviors, and conditions shift over time, so useful prediction depends on adapting without discarding everything the model has already learned.
The practical idea is simple: adaptation is not a repair after deployment. It is part of the system’s design.
That perspective continues to shape how I think about production AI—monitor drift, update selectively, evaluate longitudinally, and connect model behavior to real outcomes.
View the IEEE record ↗Research approachAn adaptation loop for
An adaptation loop for
systems in motion.
Update selectively
Move the parameters that need to change while preserving knowledge that remains useful.
Measure through time
Judge the learning process across shifts and sequences, rather than at one frozen benchmark.
Research to productionIdeas that still shape
Ideas that still shape
how I build.
Changes in data should inform product and model decisions together.
A dependable system must stay useful as conditions and behavior evolve.
The path from an outcome back to the system should be intentional.
Academic recordRead the paper
View publications →