Research Projects

Published · VLDB Endowment 2026 · ML Systems, CUDA, Distributed Training, Graph Learning

PRISM introduces a memory computation graph (MCG) built with custom CUDA kernels, where diffusion-based propagation eliminates staleness in memory-based temporal GNN training at large batch sizes. Extending to multi-node distributed training using PyTorch DDP and NCCL.

• Custom CUDA kernels for high-throughput memory propagation
• Chunk-based streaming sampler to decouple data ingestion from training
• Eliminates staleness that degrades model accuracy at scale
• Multi-node distributed training extension in progress

Ongoing · Agentic AI, Relational Deep Learning, RL, LLM

A system for answering predictive queries (churn, fraud, conversion) over relational databases without manual feature engineering, combining temporal graph learning with agentic pipelines.

• Supervised path: Heterogeneous M-TGNN encoder over temporal relational graphs for downstream prediction
• Agentic path: 5-stage pipeline — DB profiling → context graph → task policy → guided row retrieval → SLM head — replacing dedicated feature machines with policy-guided retrieval

arXiv:2607.23815 · LLM Serving, Database Systems, Inference Optimization

Kalypso is an LLM serving system designed for semantic query processing over relational data. It uses structured context batching and KV-cache reuse across queries that share relational context, significantly reducing redundant computation and improving throughput.

• Structured context batching groups queries sharing table/schema context
• KV-cache reuse eliminates redundant prefill across related queries
• Co-designed for semantic NL-to-SQL and relational question answering

Ongoing · Security, Graph Neural Networks, Anomaly Detection

Applies dynamic graph learning over OS-level provenance graphs — capturing processes, files, and sockets — to detect stealthy multi-stage intrusions using temporal dependency modeling.

• Models OS audit logs as a dynamic provenance graph with temporal edges
• Detects advanced persistent threats (APTs) via anomalous dependency patterns
• Leverages memory-based temporal GNN techniques developed in PRISM

Completed · Graph Mining, Sequential Pattern Mining

A line of work developing efficient algorithms for mining weighted and utility-based patterns from graph and sequence databases, published across PAKDD, ICDM, and Applied Intelligence.

• Graph Pattern Mining: Weighted frequent subgraph mining algorithms (WFSM-MaxPWS, WeFreS, MaxPWSCan, UGMINE) with pruning strategies outperforming baselines in runtime and scalability
• Sequential Pattern Mining: Weighted frequent sequential pattern mining (WFSPM) with support-weight formulations capturing item significance beyond frequency