Mouadh Bouali · Research Engineer
Making model behavior
measurable.
I investigate how learning systems change, fail, and recover — by modifying the mechanism, instrumenting the behavior, and letting controlled experiments decide what to keep.
Selected research
One project, followed all the way through.
T-JEPA / Collapse Dynamics
A controlled investigation of representation collapse and row-specific signal in tabular JEPA.
An open technical write-up is available for readers who want the experiment ledger, controls, and qualified results.
Working method
Hypothesis → measurement → next decision.
The useful result is rarely the first attractive curve. I build the smallest intervention that can distinguish competing explanations, then track its effect on dynamics, geometry, and downstream behavior.
That means preserving controls, recording what was retracted, and separating a faster transition from a better final representation. The aim is a model story that survives contact with its measurements.