Publications

We present STIM (Spatio-Temporal Inference Model), a transformer-based approach for predicting uncorrectable memory errors in data center servers. By analyzing spatial and temporal patterns in memory telemetry data, STIM enables proactive replacement of failing memory modules before they cause server crashes, significantly improving data center reliability.

We present BERTVision, a parameter-efficient approach that leverages BERT's hidden-state activations to achieve competitive performance on visual question answering tasks. Our model architectures are trained on BERT's intermediate representations, enabling efficient transfer learning while maintaining accuracy comparable to full-scale transformer models.

We introduce Ticket-BERT, a specialized language model for automating the classification of incident management tickets in enterprise IT environments. Our approach leverages domain-specific fine-tuning to achieve high accuracy in categorizing and routing trouble tickets, significantly reducing manual triage effort and improving response times.

We present a high-performance compression approach for transformer-based NLP models that achieves 209x reduction in parameters while maintaining accuracy comparable to full BERT models. Our method enables efficient deployment of large language models in resource-constrained environments without significant performance degradation.