High Performance Scientific Computing
Parallel algorithms, distributed workflows, MPI-based numerical methods, Dask pipelines, and batch execution on HPC systems.
University of New Mexico Computer Science
Saif Ryan Gangaram
Staff Data Scientist 2 at LANL and UNM Computer Science Ph.D. student doing active research in AI/ML, HPC, Big Data, and systems, with a focus on scalable systems and algorithms for AI/ML training and inference.
Paper accepted to IEEE BigData 2026- Current Role
- Staff Data Scientist 2, LANL
- Research Areas
- AI/ML, HPC, Big Data, Systems
- Ph.D. Program
- CS, UNM
About
Scientific computing for large, noisy, consequential data.
I am a Computer Science Ph.D. student at the University of New Mexico with a background in mathematics, data science, software engineering, and applied machine learning. I am doing active research at both Los Alamos National Laboratory and UNM in AI/ML, high performance computing, Big Data, and systems, and much of my current work centers on scalable systems and scalable algorithms for AI/ML training and inference.
I am advised by Dr. Amanda Bienz at UNM. Professionally, I currently work as a Staff Data Scientist 2 on the Test Engineering and Data Analysis team at Los Alamos National Laboratory, after previously working as a Software Engineer in Automation and AI/ML at Space Dynamics Laboratory and as a Teaching Assistant at the University at Buffalo.
I am especially interested in scalable algorithms, high-throughput data systems, time-frequency analysis, robust ML inference workflows, and the engineering practices that turn research code into dependable software. I am also building PolytopeOS and the Polytope Compiler, a from-scratch research operating system and compiler toolchain. This public site intentionally keeps project descriptions technical, high level, and suitable for open audiences.
Research
Research Themes
Scalable AI/ML Systems
Scalable systems and scalable algorithms for AI/ML training and inference: distributed training, high-throughput and low-latency inference, performance engineering, and low-level systems foundations including compilers and operating systems.
Big Data and Scalable Analytics
Current research includes Big Data methods for large scientific datasets, distributed processing, scalable storage formats, and analytics workflows for terabyte-scale data collections.
Signal Processing and Time Series Analysis
Denoising, spectral analysis, stationarity testing, transient detection, and quality checks for high-volume sensor data.
AI/ML for Scientific Data
Generative AI, supervised and unsupervised learning, clustering, regression, deep learning, anomaly detection, segmentation, forecasting, and validation for large scientific datasets.
Reliable Data Systems
Reproducible ingestion, schema design, workflow hardening, data validation, and visualization for large experimental collections.
Publications
Accepted Research
Accepted · IEEE BigData 2026
Boundary-Correct Out-of-Core Deep Inference for Scientific Sensor Archives
IEEE International Conference on Big Data, 2026
An out-of-core workflow for applying trained deep models to scientific sensor archives. Halo-aware slabs, aligned time-frequency transforms, and manifest checks preserve results at chunk boundaries while keeping memory use bounded. On a 16-channel trace, slabbed processing matched every detection mask from the whole-file pass.
Accepted for publication. The official IEEE link will be added when available.
Selected Work
Projects
Protecting Tail Latency Under Contention
A team experiment comparing FIFO, FQ-CoDel, and CAKE when small requests share a bottleneck with bulk TCP traffic. I lead the architecture, foreground workload, trial runner, and integration. The tooling includes isolated Linux namespaces, paced workloads, and independent run validation. The study is ongoing; primary results are pending.
View open-source repositorySignal Foundry
Brings AlphaForge and Signalattice into one research workspace while preserving their histories and separate environments. A local control service runs bounded jobs, and the Nexus browser interface configures, compares, and inspects results. The project supports research and explicitly authorized paper operations; live trading is not implemented.
View open-source repositoryHyperion Z
A self-hosted console for coordinating model-powered development tasks. Its Autopilot plans dependent tasks, runs tool-using sessions, and presents the resulting Git diff. The browser also provides streaming output, a Monaco editor, tmux control, saved SSH connections, and persistent agent configurations.
View open-source repositoryCrucible
A local training-data platform with batch and streaming ingestion, quality checks, exact and near-duplicate removal, versioned datasets, and deterministic training shards. Its research harness measures how data quality, mixing, deduplication, and scale affect a small transformer. The reference workflow runs offline on a laptop CPU.
View open-source repositoryInternal Agentic AI Workflow Harness
Built a Codex and Claude Code-style internal agentic AI harness from scratch for approved enterprise LLM APIs, with multi-agent orchestration, software-development assistance, collaboration context retrieval, meeting summarization, email and workflow drafting, and document generation for controlled internal environments.
SynaptiCast
A neural telemetry research platform for response decoding and causal signal forecasting. The development benchmark uses simulated subjects, held-out-subject evaluation, calibrated uncertainty, and corruption tests, with classical baselines alongside neural models. Results describe software and simulated signals, not clinical or implanted-device performance.
View open-source repositoryAxiomStack / DorkLLM
A local language-model lab connecting the from-scratch DorkLLM decoder, evaluation and document retrieval, and the DorkChat streaming interface. The development platform includes causal attention, KV caching, instruction tuning, typed APIs, and browser tests. Models and reference experiments are deliberately small enough to inspect and run locally.
View open-source repositorySignalattice
Probabilistic market forecasting with chronological evaluation, calibration, causal sequence models, and cost and delay analysis. Recent work adds a versioned DuckDB/Parquet feature store, resumable research jobs, and a read-only local evidence API. Each report records why a forecast passes or fails its declared decision criteria; simulations do not establish live profitability.
View open-source repositoryAlphaForge
Quantitative ML research with purged temporal validation, causal regime models, costed backtests, and a C++ order-book simulator. Recent work adds constrained Markowitz optimization with independent optimality checks, factor-risk attribution, and reproducible model-selection records. The research retains rejected candidates and adverse results; no candidate is qualified for live trading.
View open-source repositoryLearning Systems Atlas
A documented v1.0 research release spanning from-scratch supervised and unsupervised models, NumPy autodiff, PyTorch training, and reinforcement learning with REINFORCE, DQN, and PPO. Reproducible studies include held-out evaluation, model cards, diagnostic plots, and comparisons with scikit-learn that preserve both agreement and disagreement.
View open-source repositoryConductor
Developing an agentic AI dashboard for software development and workflow optimization, with a focus on coordinating tasks, surfacing project state, supporting debugging workflows, and improving how developers move from intent to verified changes.
PolytopeOS
A research OS and compiler with a reproducible x86_64 UEFI boot path: a Rust loader validates and launches a separate freestanding kernel, which checks the boot contract and emits structured diagnostics. The project also includes a compiler lexer and developer CLI. Current execution targets headless QEMU; broader resource-management features remain research goals.
View open-source repositoryParallel Principal Component Analysis
Implemented distributed PCA with rank-local data sharding, collective covariance aggregation, eigenpair broadcast, optional standardization, and scalability logging for large datasets.
Large-File Inference Workflow
Boundary-correct processing for scientific sensor archives using halo-aware slabs, bounded PyTorch inference, and validated NetCDF products. The associated paper has been accepted to IEEE BigData 2026 and checks slabbed outputs against whole-file processing.
Read the publication summaryReal-Time Emotion Recognition
Trained and evaluated CNN-based models using large public image datasets, applying augmentation, regularization, feature extraction, and real-time video processing techniques.
Financial Risk Modeling
Developed supervised learning pipelines for probability-of-default estimation with feature engineering, hyperparameter tuning, stress testing, and error analysis.
Scientific Data Access Tools
Designed database-backed analysis tools for large scientific collections, with searchable metadata, plotting, export workflows, and reproducible input generation for modeling studies.
Segmentation and Detection Models
Applied CNNs, U-Net style segmentation, clustering, and classical image processing to improve detection and denoising in scientific imagery and signal-derived products.
Reproducible Pipeline Hardening
Built instrumentation, validation checks, runtime tracing, and portable execution paths for scientific data workflows, improving reproducibility across local, Linux, and HPC-style environments.
Background
Experience and Education
2026 - Present
Staff Data Scientist 2, Los Alamos National Laboratory
Research data scientist on the Test Engineering and Data Analysis team. Develops scalable acquisition, analysis, and compression workflows, writes computer science research papers, and builds HPC training and inference pipelines for scientific datasets totaling more than 700 TB. Work also includes signal processing, statistical validation, large-codebase integration, and full-stack agentic AI tools with human review.
2023 - 2026
Software Engineer, Space Dynamics Laboratory
Built data pipelines, database-backed web tools, scientific analysis workflows, and automation software for aerospace and remote-sensing research settings.
2024 - 2025
Technical Recognition and Presentation
Received internal recognition for user-centered engineering, data tooling, documentation, and technical communication; presented telemetry anomaly-detection methods at a public technical venue.
2019 - 2020
Teaching Assistant, University at Buffalo
Supported lab and recitation instruction, grading, exam administration, and student mentoring for undergraduate computer science coursework.
University of New Mexico
Ph.D. in Computer Science, expected 2028; advisor: Dr. Amanda Bienz
University at Buffalo
M.P.S. in Data Sciences and Applications, 2023
University at Buffalo
B.A. in Mathematics, Computing and Applied Mathematics, 2021
Methods
Technical Areas
Python
C++
C#
SQL
MATLAB
Julia
PyTorch
TensorFlow
Scikit-learn
Codex
Claude Code
LLM APIs
Agentic AI
Multi-Agent Systems
RAG
LLM Evaluation
FastAPI
Streamlit
Quant ML
Backtesting
Portfolio Construction
Risk Analytics
NumPy
Pandas
Dask
Slurm
MPI
PostgreSQL
SQLite
NetCDF
HDF5
Parquet
DuckDB
Data Quality
Dataset Versioning
Deduplication
Docker
Linux
Git
Time Series
Spectral Analysis
Visualization
Contact
Open to research conversations and technical collaboration.
For public research, academic, or professional communication, email is the best way to reach me.
Public Information Note
The descriptions on this site are intentionally limited to public-safe topics. They omit sensitive access details, restricted system names, operational parameters, non-public datasets, and program-specific details.