Research

• Evolutionary tree is being used in many different applications, from understanding the evolutionary history of species to identifying genetic variants associated with diseases. Inferring species trees from raw genome assemblies is normally a slow, multi-stage process.
• I led the development of ROADIES which is an end-to-end automated and scalable framework to construct species trees directly from whole genomes. ROADIES automates workflow orchestration end to end and parallelizes sequence alignment and gene-tree inference across CPU and GPU resources.
• The work was published in PNAS (2025), selected as the issue's journal cover, and presented at ISMB 2024. It is open source and being used by researchers as part of the Vertebrate Genomes Project (VGP) consortium and the Scripps Institution of Oceanography at UC San Diego.

• Dynamic-programming algorithms underlie many bioinformatics pipelines — sequence alignment, homology search, basecalling — but hand-coding an FPGA accelerator for each one is slow and requires hardware design expertise.
• DP-HLS is a reusable High-Level Synthesis (HLS) framework with a parameterizable accelerator architecture and integrated HLS directives, letting algorithm developers target FPGAs without optimizing each kernel manually.
• We implemented 15 diverse DP kernels as example and deployed them on AWS EC2 F1 FPGA instances. The work was published at IEEE HPCA 2026.

• As genomic databases are growing massively, regenerating species tree from scratch is computationally wasteful.
• ROADIES-XP updates anexisting species tree by placing newer genomes onto it, instead of regenerating it de novo. It accelerates the underlying compute intensive operations on GPUs.

• Many genomic datasets combine fragmented or heterogeneous sequence data, and there is no standard way to integrate all types of sequence data into a single analysis.
• MAGNET is ongoing research into integrating this kind of fragmented or heterogeneous sequence data with a genome-based phylogenetic backbone, combining scalable locus identification, alignment, and gene-tree inference with an existing species-tree pipeline. The project is still under active development.

Parallelized Suffix Array construction on GPUs, achieving an 86–571× speedup over optimized CPU baselines for genomic read mapping. Also accelerated dense matrix multiplication using CUDA on NVIDIA K80/T4 GPUs with Intel AVX2 vectorization, and optimized large-scale solvers (Aliev-Panfilov) on the Expanse Supercomputer using MPI and C++.

Improved the HLS4ML library for efficient ML hardware inference via High-Level Synthesis. Partnered with CERN's HLS4ML team to implement feature enhancements and optimizations on the DL-to-FPGA flow.