PRISM: A Training System for Temporal Graph Learning Through Staleness Avoidance
Custom CUDA kernels build a memory computation graph (MCG); diffusion-based propagation eliminates staleness in M-TGNN training at large batch sizes.
Md Ashraful Islam
PhD Candidate · UMass Amherst · ML & Data Systems
I am a PhD candidate in the Computer Science Department at UMass Amherst, advised by Prof. Marco Serafini. My research focuses on ML & Data Systems, Agentic AI, and Relational Deep Learning. I build high-performance systems with CUDA, PyTorch DDP, and NCCL for large-scale temporal and relational graph learning (VLDB’26). Co-author of Kalypso (relational LLM serving, arXiv’26).
News
- 2026 Research Internship at Tursio Inc. (Bellevue, WA) — built a context-graph enrichment pipeline for an agentic NL-to-SQL system.
- 2026 Preprint on arXiv: Kalypso: Relational LLM Serving — co-authored with Hojae Son and Marco Serafini.
- 2026 Paper accepted in VLDB. PRISM: A Training System to Unlock the Potential of Temporal Graph Learning Through Staleness Avoidance
- 2025 Paper accepted in International Journal of Speech Technology. STFT-GradTTS: A Robust, Diffusion-based Speech Synthesis System with iSTFT decoder for Bangla
- 2024 Paper accepted in Applied Intelligence. Graph-based substructure pattern mining with edge-weight
- 2023 Started PhD at UMass Amherst on ML System
- 2023 Paper accepted in Applied Intelligence. UGMINE: utility-based graph mining
- 2022 Promoted to Assistant Professor (University of Dhaka, Dept. of CSE)
- 2022 Paper accepted in Applied Intelligence. Weighted frequent sequential pattern mining
- 2019 Started new position as Lecturer (University of Dhaka, Dept. of CSE)
- 2018 Paper accepted in PAKDD. WFSM-MaxPWS: An Efficient Approach for Mining Weighted Frequent Subgraphs from Edge-Weighted Graph Databases
Publications
Kalypso: Relational LLM Serving
LLM serving system for semantic query processing: structured context batching and KV-cache reuse across queries sharing relational context.
STFT-GradTTS: A Robust, Diffusion-based Speech Synthesis System with iSTFT Decoder for Bangla
Graph-based Substructure Pattern Mining with Edge-Weight
Software
PRISM
Open-source framework for staleness-free and scalable temporal GNN training. Features streaming samplers, chunk-based memory graphs, and optimized CUDA kernels for high GPU utilization.
Kalypso
Relational LLM serving system for semantic query processing: structured context batching and KV-cache reuse across queries sharing relational context.
WFSPM
Implementation of Weighted Frequent Sequential Pattern Mining — a framework for mining weighted sequential patterns efficiently using pruning and weighting strategies.
MaxPWSCan
Implementation of Graph-based Substructure Pattern Mining with Edge-Weight — a graph mining algorithm that discovers weighted frequent substructures using MaxPWS pruning.
STFT-GradTTS
A diffusion-based Bangla text-to-speech (TTS) system featuring a Stochastic Duration Predictor and a multi-stream iSTFT decoder for high-quality waveform generation.
Skills
Languages: Python, C++, C, Java, SQL, JavaScript
ML/Graph Frameworks: PyTorch, PyG, DGL, CUDA (custom kernels), PyTorch DDP, NCCL
Databases & Systems: PostgreSQL, SQLite, Linux, Git, Docker
Profiling & Optimization: PyTorch Profiler, nvprof, perf
Experience
-
Research Intern, Tursio Inc. — Bellevue, WA
Jun – Aug 2026
Built context-graph enrichment pipeline for an agentic NL-to-SQL system; added LLM-in-the-loop result interpretation producing provenance-linked explanations. -
Assistant Professor, University of Dhaka — Dhaka, Bangladesh
Sep 2019 – Aug 2023
Taught Algorithms, Database Systems, ML (30–60 students/class); supervised 6+ MSc/undergrad theses; co-PI on two funded projects. -
Lecturer, East West University — Dhaka, Bangladesh
Apr 2018 – Aug 2019
Taught Data Structures, Algorithms, and Discrete Mathematics.
Teaching
- CS 589 – Machine Learning, TA, UMass Amherst (Fall 2025)
- CS 220 – Programming Methodologies, TA, UMass Amherst (Fall 2023, Spring 2024)
- CS 3203 – Algorithms II, Instructor, University of Dhaka (2021, 2022)
Service & Talks
Service
- Reviewer: STI (2020, 2021, 2022)
- Undergraduate Research Mentor (Winter 2024)
- ERSP PhD Mentor (Fall 2023, Spring 2024)
Talks
- "Challenges of Scalable Memory-based Temporal Graph Learning" — UMass ML Systems Reading Group, 2024
- "PRISM: Unlocking True Potential of Memory-based Temporal Graph Learning" — Reading Group, 2024
Contact
Email: mdashrafulis@umass.edu
Office: LGRC A249, University of Massachusetts Amherst
Social: GitHub · Google Scholar · Semantic Scholar · ORCID