Unsupervised Learning
Variational Autoencoder
Model Distillation
Boosted Decision Trees
Real-Time ML
I designed and deployed an unsupervised anomaly-detection system for more than one billion high-frequency events. The system was validated against independent signal samples and deployed for continuous use.
Results
- Recovered 4.5–21% additional rare events missed by existing threshold-based methods
- Distilled a variational autoencoder into a boosted-decision-tree model
- Reduced inference time to under 3 microseconds
- Adapted the model for limited memory and computing resources
- Built validation and monitoring workflows for continuous operation
Publication
Joint first author of “Chopping and distilling variational autoencoders for real-time anomaly detection in high energy physics”, arXiv:2606.14857.
Tech Stack
Python, PyTorch, variational autoencoders, boosted decision trees, model distillation and hardware-aware inference.