AI & ML interests
We build tools to understand how models change during training, identify where regressions and unwanted behaviors emerge, localize meaningful changes within the model, and correct or remove learned behavior without full retraining.Our work spans training dynamics, model interpretability, machine unlearning, training-free model optimization, and AI governance.
Recent Activity
View all activity
Articles
Sequential fine-tuning quietly degraded a capability the model already had. Authentrics restores it — no retraining from scratch.
Reading how Mixtral-8x7B routes activations, and how that structure emerges across training — without moving weights off your hardware.
A BadNets backdoor is planted in one MNIST training epoch, localized via checkpoint drift, and removed with Authentrics — no retraining from scratch.
A model fine-tuned on data containing PII memorized it. Authentrics locates and removes that influence — no retraining from scratch.
Tracking parameter and activation drift across the NVIDIA Nemotron Cascade 8B checkpoint lineage to see what changed, where, and when.
Models used for testing Authentrics.ai software suite
Sequential fine-tuning quietly degraded a capability the model already had. Authentrics restores it — no retraining from scratch.
A model fine-tuned on data containing PII memorized it. Authentrics locates and removes that influence — no retraining from scratch.
Reading how Mixtral-8x7B routes activations, and how that structure emerges across training — without moving weights off your hardware.
Tracking parameter and activation drift across the NVIDIA Nemotron Cascade 8B checkpoint lineage to see what changed, where, and when.
A BadNets backdoor is planted in one MNIST training epoch, localized via checkpoint drift, and removed with Authentrics — no retraining from scratch.
Models used for testing Authentrics.ai software suite