This summarizes some main highlights and results from my work since 2000, most recent first, from projects and information that can be shared publicly.
It is divided into
- Geospatial Data Science (2025-present)
- Generalist Data Science (2013-25)
- Astrophysics (2000-13)
I am grateful to my many colleagues without whom these highlights would not have been possible.
Table of Contents
Geospatial Data Science (2025-present)
This section is necessarily short due to the recency of my change to this field.
An initial highlight is that personal projects resulting from my transition to this field have worked well, including the Don Valley Tour, Millbrae’s Steepest Streets, and Countries.
These projects use QGIS, PyQGIS, Python, UMap, OpenStreetMap, OSMNx, Dash-Leaflet, PostGIS, GeoPandas, Pytorch, TorchGeo, and others.
In future, I expect the Universe Maps to be a personal project highlight, alongside freelance professional projects in data science, GIS, and GeoAI.
Generalist Data Science (2013-25)
LLMs on multinode H100s: Pretraining and Finetuning MosaicML Models (2024)
Acquired by DigitalOcean (DO) in 2023, Paperspace’s expertise in GPUs and AI formed the foundation of DigitalOcean’s successful move into the AI space from 2024 onwards.
Part of this success required showing that real large language models (LLMs) could be used end-to-end on DO Nvidia GPU infrastructure. MosaicML models were well suited to this because they allowed pretraining, finetuning, and inference end-to-end on real large datasets using tools and infrastructure suited to production. Multi-GPU and multinode machines allowed us to perform these steps on models of up to 30 billion parameters. Comparison to MLPerf benchmarks showed that we were running at state-of-the-art performance.
Paperspace Firsts: Workflows, LLM finetuning, Multinode (2020-23)
Similar to Digital Ocean, as a general cloud platform providing GPU compute to the public while having a core value proposition of having better service and being easier to use than the hyperscalers (Amazon Web Services, Google Cloud, Microsoft Azure), Paperspace required showing that real end-to-end data science could be done on its platform. This entailed a number of first-time showings including Gradient Workflows, finetuning of full-size large language models, and multinode Nvidia H100 GPUs.
Paperspace End-to-End Data Science: Nvidia Merlin (2023)
Showing end-to-end functionality produced a variety of technical content as a byproduct. The example of running Nvidia Merlin recommender systems end-to-end used ETL/ELT at scale, feature engineering, a feature store, synthetic data, model training & saving, deployment, and model ensembles (deployment business logic). Tools used included Nvidia Merlin NVTabular (larger than memory datasets), Nvidia RAPIDS cuDF (large dataframes), Dask, Apache Parquet, Feast, and Nvidia Triton Inference Server, with everything working correctly.
Other Paperspace project-based content included Mosaic MPT & Llama 2 LLM finetuning, Stable Diffusion with self-attention guidance, gradient-boosted decision trees, generalist models (image+text+audio+video), Graphcore IPUs, PyTorch Lightning, Hugging Face Accelerate, decision transformers, TensorFlow Distribution Strategies, and using Gradient Workflows in production to update our public datasets.
Driving Dotscience MLOps Platform Product Roadmap, Developer Experience & Content Strategy (2019−20)
As the most senior data scientist in this early stage startup company (named Principal but de facto Chief Data Scientist), I owned and drove the product roadmap for Dotscience, reporting directly to the CEO. In this role I integrated feedback from sales, marketing and engineering as well as my own professional opinion, skills and experience into defining and prioritizing the most critical revenue−driving features, such as RBAC, A/B testing and enhanced statistical model monitoring capabilities. I also took ownership of the product documentation to drive improved developer experience on the platform, as well as the content strategy in which I authored key blog posts that staked out the company’s whole value proposition in a way that made sense to the target audience such as Why do Data Scientists Need DevOps for Machine Learning (MLOps)?
Previous roadmap areas I have covered include auto-ml, auto-featurization, missing values, model performance, data preparation, time series, scalability, as well as many smaller areas and details from the whole end-to-end data science proess.
NYC Taxi Blog (2016)
I wrote various blog entries for Skytree. The New York Taxi data one was nice because we showed a 500 million row dataset being used as the training set for a machine learning model, fully within the Skytree graphical user interface. It also included data preparation and featurization from the raw public data.
Trillion Element Training Set (2016)
I was part of the Skytree data science and engineering teams that showed its command line using a dataset of 10 billion rows by 100 columns (hence 1 trillion elements) as the training set for its distributed gradient boosted decision tree on 10 Hadoop nodes. This was several years before modern LLM training sets, and there were no comparably sized datasets described publicly at the time.
Skytree Benchmarking (2015)
I led the benchmarking project of Skytree versus various competitor software, showing single node CPU speedups of 150x vs. R, 100x vs. Scikit−learn, 71x vs. Spark MLlib, and 2x vs. H2O and XGBoost. Multi-node then increased these further via both weak and strong scaling. These results were used extensively in the company’s communications with customers and investors.
Skytree PoCs (2013−17)
At Skytree we undertook customer projects across a wide variety of industry domains as PoCs, solving business problems. A PoC (proof−of−concept) has been used by many companies looking to add AI to their business as a way of showing that it can generate business value for them. Skytree never lost a proof-of-concept (PoC) customer engagement to a rival company based on performance. I led or participated in PoCs with over 60 companies, including deal sizes of $1m+, international and onsite work.
Skytree was written from the ground up in C++, which was the primary reason for its speed, in combination with the expertise of engineers from Alex Gray’s Georgia Tech FASTlab, from which the company was founded.
Astrophysics (2000-13)
CANFAR+Skytree (2010−13)
I initiated and led a collaboration between CANFAR, a project of the Canadian Astronomy Data Centre where I worked from 2009−13, and Skytree “The Machine Learning Company” founded in 2010. CANFAR provided a cloud-like system for astronomers where they could access large compute resources via virtual machines, but with the queuing functionality of a supercomputer. Adding Skytree to it thus opened up the C++ machine learning speed highlighted in Skytree Benchmarking above, combined with the possibility to run it in parallel across 500 nodes. This was one of the first systems in the world to open up large scale machine learning (ML) to astronomy in this way. It was a precursor to the larger scale in which ML is used in astronomy today.
IJMPD Review: Data Mining and Machine Learning in Astronomy (2009−10)
In 2009 I was invited to write a peer-reviewed literature review on Data Mining and Machine Learning in Astronomy for the International Journal of Modern Physics D. The resulting 61 page article published in 2010 remains cited to this day (Google Scholar). Similar to CANFAR+Skytree’s pioneering infrastructure, it was one of the first extensive published articles on this subject. Data-intensive work and use of machine learning are widespread in astronomy today.
Quasar Photometric Redshift Probability Density Functions (2008)
We used machine learning to produce a sample of quasars with all accurate distances, without a fraction of catastrophic failures where a significant number of the distances are completely wrong. This had not previously been shown in the literature.
Quasars are the bright centers of some galaxies, and photometric redshifts refer to measuring the distances to astronomical objects using only their images without needing to take a spectrum. This is useful because there are far more objects with images than spectra. Data about the universe’s structure is enriched, and this is valuable for many different science questions. Photometric redshift distances can be assigned using machine learning with objects that have spectra as a ground-truth training set.
Catastrophic failures in the redshifts come about for quasars because at certain distances the bright lines in the spectra used for the training set drop between the color filters on a telescope. This means that quasars at different distances can appear the same color, making their distance ambiguous.
Our machine learning solved this by using k nearest neighbors, with the counterintuitive setting of k=1. The failures stopped blending into the good results and became their own area on the diagram of predicted distance versus true distance. By itself this is not useful because you would still need a ground truth, but when we perturbed the inputs we could generate a “probability density function” (PDF) of possible distance values for each object. The objects that were failures had two peaks. Removing the objects with 2 peaks left a subsample that had accurate distances.
Morphological Galaxy Classification Using Artificial Neural Networks (2000−04)
Forming my masters thesis in 2000 and part of my PhD work from 2001−04, this included deep learning in the form of artificial neural networks (ANNs) with more than one hidden layer, using the classical fully-connected dense architecture with backpropagation. We showed that ANNs could classify galaxies in the Sloan Digital Sky Survey (SDSS) into morphological types such as Hubble classes with the same accuracy as human experts, the first study to do so on this large digital sky survey.
The later well−known Galaxy Zoo project (which I was not part of) collected a few hundred thousand classifications by crowdsourcing for some years, whereas our ANNs could assign classifications at a rate of this many per minute. Galaxy Zoo produced many other results, and a combination of human input and machine-assigned object classification is widely used in astronomy today on the SDSS, its successors, and many other large datasets.
