Software

Most of my research ships as a PyTorch library rather than a one-off script. tslens is a Captum-compatible interpretability toolkit that attributes any callable mapping (batch, seq_len, n_features) to predictions, with no registry to subclass and nothing to reimplement per model: 14 attribution methods, tested against 30+ time series architectures from linear models to LLM-backed foundation models. AstroSpec, AstroGen and AstroLens take the opposite approach where it fits better, putting published astronomy models behind one registry and interface while preserving survey-specific calibration and native instrument grids instead of resampling everything to a common shape.

The work behind them is as much systems as modeling: distributed data-parallel training on SLURM clusters, torch.compile and mixed precision for 3D volumes that do not fit on one GPU, inference optimization for serverless deployment, and GitHub Actions pipelines that test and publish the packages.

Stack

Core: PyTorchPyTorch LightningNumPyPandasScikit-learnTensorFlow
Models: Hugging Face TransformersDiffusersAcceleratetimmCaptum
Scale: Distributed Training (DDP)Multi-GPUMixed PrecisionSLURM / HPCInference Optimization
Cloud: AWS LambdaSageMakerS3DockerServerless Deployment
Engineering: GitGitHub Actions CI/CDPython Packaging (PyPI)Experiment Tracking (W&B)Technical Documentation
Languages: PythonC++SQLSwiftBash

Libraries

tslens 13

Time series interpretability for PyTorch: 14 attribution methods behind one Captum-compatible interface, tested against 30+ architectures from linear models to LLM-backed foundation models. Published on PyPI with a MkDocs reference site, pytest suite and GitHub Actions release pipeline.

GitHub PyPI Docs

PyTorchCaptumExplainable AITime-Series MLPublic API DesignpytestGitHub Actions CI/CDPyPI PackagingTechnical Documentation

AstroSpec

Unified spectral library for astronomy, putting seven spectroscopic models behind one PyTorch interface for source classification, redshift estimation, stellar property inference and self-supervised representation learning on native instrument grids.

GitHub

PyTorchAstronomical SpectroscopyFoundation ModelsSelf-Supervised LearningRepresentation LearningTransformersDistributed TrainingMulti-GPU

AstroGen

Generative models for astronomical data in 1D, 2D and 3D, covering diffusion models and VAEs for super-resolution, denoising and conditional generation over images, spectra and simulations.

GitHub

PyTorchGenerative AIDiffusion ModelsVariational AutoencodersComputer VisionHugging Face Diffusers3D Deep LearningDistributed Training

AstroLens

Unified library of vision models for astronomy, covering galaxy morphology classification, strong gravitational lensing and multimodal representation learning, with pretrained foundations such as AstroCLIP and AION-1.

GitHub

PyTorchComputer VisionVision TransformersMultimodal LearningTransfer LearningPretrained ModelsDistributed Training

Paper Code

Cosmo3DFlow

KDD 2026

Wavelet flow matching for reconstructing the early universe from cosmological simulations.

Code Paper

PyTorchPyTorch LightningFlow MatchingWavelet Transforms3D U-NetScientific Machine LearningMulti-GPU TrainingSLURM / HPC

AI for Astronomy

IJHPCA 2025

Cloud framework for data-parallel astronomical model inference on AWS serverless infrastructure.

Code Paper

PyTorchtimmModel DeploymentAWS LambdaAWS Step FunctionsAmazon S3DockerServerless ComputingDistributed Inference

Timeseries-Explained

AAAI 2025 Workshop (AI4TS)

Reference implementation of WinTSR, a windowed temporal saliency rescaling method, benchmarked against local interpretation baselines.

Code Paper

PyTorchExplainable AITime-Series MLModel EvaluationBenchmarkingReproducible Research

Financial Time Series 19

IEEE BigData 2024 Workshop (LLMs for Finance)

Benchmarks transformer and LLM-based forecasters (PatchTST, iTransformer, TimesNet, GPT4TS, TimeLLM) on stock, commodity and financial aid data.

Code Paper

PyTorchLarge Language ModelsFine-TuningHugging Face TransformersTime-Series ForecastingModel EvaluationBenchmarking

COVID-19 Age Groups

AAAI 2024 Workshop (AI4TS)

Interpreting the impact of population age groups on COVID-19 infections with deep learning time series models.

Code Paper

PyTorchTransformersExplainable AITime-Series ForecastingPublic-Health AnalyticsModel Evaluation

SA-Timeseries

AAAI/SIGAI Doctoral Consortium 2024

Sensitivity analysis of temporal dependencies and spatio-temporal patterns in time series models.

Code Paper

PyTorchSensitivity AnalysisExplainable AITime-Series ForecastingSpatiotemporal ModelingBenchmarking

GPCE-COVID

IEEE ICDH 2023

Interpreting county-level COVID-19 infections using transformer and deep learning time series models.

Code Paper

PyTorchTransformersTime-Series ForecastingPublic-Health AnalyticsDockerReproducible Research

MVAM

CPS-IoT Week 2023

Multi-variant memory attacks on IoT trust computing.

Code Paper

CybersecurityIoT SecurityEmbedded SystemsTrust ComputingMemory Security

DP on NLP Bias

AAAI 2023 Workshop (PPAI) / Data Engineering Bulletin 2024

Measuring how differential privacy during fine-tuning affects bias in pretrained language models.

Code Paper

PyTorchNatural Language ProcessingHugging Face TransformersFine-TuningDifferential PrivacyResponsible AIBias EvaluationModel Evaluation

Predict Code Changes

Information and Software Technology 2022

Early prediction of merged vs abandoned code changes in modern code review.

Code Paper

PythonApplied MLPredictive ModelingFeature EngineeringClassification

Anomaly Detection on UNSW-NB15 9

ITNAC 2020

Network traffic anomaly detection with LightGBM and ensemble gradient boosting on the UNSW-NB15 benchmark.

Code Paper

PythonLightGBMGradient BoostingAnomaly DetectionFeature EngineeringApplied ML

Other Projects

VLASS Vision 5

Vision models for classifying radio astronomy images from the VLA Sky Survey, comparing CNN and transformer backbones (ResNet, MobileNet, ViT, SwinViT).

GitHub

PyTorchPyTorch LightningComputer VisiontimmTransfer LearningModel Evaluation

See all on GitHub →