Projects
A selection of research, professional, and personal software projects. More of my open-source work is on GitHub.
Research & Professional Projects
APSO — Adversarial Particle Swarm Optimization Cornell
An architecture-independent Explainable AI (XAI) and adversarial robustness framework for trustworthy scientific reasoning, applied to sequence-free enhancer prediction and chromatin architecture identification. Published in Bioinformatics (2025). [DOI: 10.1093/bioinformatics/btaf371](https://doi.org/10.1093/bioinformatics/btaf371)
Walrus Inference Optimization Flatiron / Polymathic AI
Automated optimization pipelines for the 1.28B-parameter Polymathic Walrus multimodal scientific foundation model — a 47% reduction in peak GPU memory and 2.3× sample throughput with accuracy within 0.2% of the FP32 baseline.
Multi-Agent Intent Resolution Siemens
A distributed recursive task decomposition system for multi-agent reasoning: intents decomposed into Directed Acyclic Graphs (DAGs) with per-path blackboards across heterogeneous LLMs, integrated with MCP-conformant microservices — up to 36× faster resolution and 25× token reduction with 100% task correctness.
Energy-Efficient Neural Network Pruning NREL / U. Iowa
Compression pipelines using Iterative Magnitude Pruning (IMP) and activation variance, compressing neural networks by up to 90% while preserving accuracy. Validated on a benchmarking framework processing a 1.7 TB dataset across 483,000 configurations and 11M training runs. See Publications.
Cell-Type Invariant Enhancer Prediction Cornell
Scalable deep learning pipelines over epigenomic chromatin datasets for genome-wide enhancer identification, mapping latent-space representations to underlying systemic biological features.
Personal Projects
Local Deployment of Agentic AI RAG Systems
Privacy-preserving retrieval-augmented generation (RAG) pipelines using open-source LLMs — semantic search, document indexing, and agent-based orchestration for research workflows.
Reproducible ML Benchmarking Tooling
Modular tooling to standardize ML experiment configuration, logging, and evaluation, enabling reproducible comparison of model architectures, pruning strategies, and training regimes. Used across collaborative research projects.
Earlier Work
Pyaesar — Multi-Node Multi-Processor API LANL
An embarrassingly-parallel HPC Python API, presented at the SciPy 2020 Conference. Proceedings
Beryllium Strength Calibration via PSO LANL
Extended the Beryllium strength model to high strain-rate regimes using particle swarm calibration, with visualization tooling; presented at ASME 2019. Technical report
Forecasting Dengue in Brazil LANL
Time-series disease forecasting with parallel computation; presented at the LANL Student Symposium, 2018.
MAMA Plugin for Nuclear Material Analysis LANL
A plugin for particle analysis in the MAMA software, developed during the Domestic Nuclear Detection Office internship, 2017.