I am a senior software engineer focused on machine learning. I have played a critical role in two paradigm shifts in fields of science using machine learning — recently, in weather prediction; previously, in materials science.

Sabbatical Projects

Cubed is a library for distributed, serverless, memory-bounded computation on the Python Array API.

• Investigating how Cubed intersects with accelerators, primarily via JAX.
• Added an affordance for compiling Cubed operations via Numba or JAX (JIT or AOT).
• Adding support for JAX arrays on M1+ hardware via jax-metal.

EarthRanger is a project funded by the Allen Institute of AI (AI2) to support African Elephant Conservation, among other conservation projects. It’s a collaborative effort involving the Mara Elephant Project and Ecoscope.io.

• dask-ee. Google Earth Engine Feature Collections via Dask Dataframes. Featured at SciPy 2024.
• xarray-sql. An experiment to join raster, vector, and point data by treating pixels as tables.

Google Experience

Anthromet is on a mission to make weather information universally accessible and useful by developing state-of-the-art AI weather forecasts and integrating them into products.

• Xee: An Xarray backend for Google Earth Engine. (Xarray, Google Earth Engine)
  • Created this package, launched as a keynote feature of Geo for Good 2023.
  • This connects Google Earth Engine to the scientific Python ecosystem.
  • When integarted with Xarray-Beam, it only takes ~25 lines of code and a few hours to export 20 TiBs of data from Google Earth Engine to Zarr, saving thousands of LOC and days of debugging quota limits.
  • Built to serve an internal weather research platform to build and ship new weather models.
  • Since launch, the project has at .
• weather-tools, a set of data pipelines to make weather data universally accessible and useful. Originally a side project (20% time), I grew the project to a team of 8 engineers to serve ~25 research and product teams across Google AI, Brain, X, DeepMind and Cloud. (Apache Beam, Xarray, Google Earth Engine, Google BigQuery, MetView):
  • GraphCast: Enabled DeepMind to ingest and regrid ERA5, the dataset behind their autoregressive graph neural network. At the time of publishing, this was the SOTA 10 day weather forecast, beating physics based models. GraphCast was among 2023's top-ten biggest breakthroughs published in Science, ushering in a new generation of AI-based weather forecasts.
  • MetNet v3 & Nowcasting in Google Search: MetNet is the world’s leading Nowcast, or 24 hour, minute by minute weather forecast at 1-4 km resolution. weather-tools created gobal training, inference, and validation datasets, enabling our team to ship to GSRP.
  • Project Contrails. I provided critical weather data and data engineering pipelines that made this project possible. This project alone will solve 1% of anthropogenic climate change by reducing solar irradiance from airplanes.
  • ARCO-ERA5 & Weatherbench2: I ingested and published the two biggest datasets in Google Cloud’s Public Dataset program. I worked with Cloud to shape weather tools to ingest ERA5 into BigQuery.
  • DeepMind’s Wind Energy optimization. By ingesting weather data into Google BigQuery, I helped create an ML model with DeepMind and Cloud to make wind energy more profitable in the Texas energy grid. This lead to a ~3% improvement of mean absolute zero error and ~$7 million more in revenue over 8 months from wind power.
• Weatherbench2. The definitive benchmark to fairly compare AI-based, mid-range weather forecasts and a cornerstone for all future ML weather model development. (Xarray, Apache Beam, Zarr)
  • In addition to ingesting the fundamental datasets (see above), I made core updates to Xarray-Beam, the underlying engine behind the benchmark.
  • Contributed code to the benchmark itself, helping ship it to production.
• ARCO-ERA5 & Pangeo Forge. (Apache Beam, Xarray, Dask, Zarr)
  • ARCO-ERA5 is the biggest dataset in Cloud Public Datasets, at 12+ petabytes. It represents the most accurate history of weather on Earth from 1940 to the present.
  • Pangeo Forge is aiming to become the canonical open ecosystem of data engineering recipes for producing cloud-optimized, analysis ready data, i.e. the conda-forge of scientific datasets.
  • I contributed bug fixes and the Beam integration upstream to Pangeo Forge to produce ARCO-ERA5.
  • I was the impetus for Pangeo Forge to transition their data engineering system to Apache Beam.
  • To support Pangeo Forge, I contributed a Dask runner to Apache Beam (and presented it at PyData).
  • Pangeo Forge would go on to cloud optimize over 4,000 CMIP6 datasets via Apache Beam.
  • The ARCO-ERA5 corpus likely includes the biggest single Zarr ever created.
• Awards and recognition received while on Anthromet:
  • Google Research’s Science Award for Best Collaboration, along with my team and partner teams.
  • A Greenie award from Anthropocene, an internal grassroots organization focused on climate technologies.
  • A spot bonus from John Platt, the head of Google Applied Sciences, for speeding up a critical data ingestion workflow by a few orders of magnitude.
• Lead a team of twelve 20%-engineers to make contributions to the weather community, including internal changes to Google Earth Engine and open source changes to Weather Tools.

Arcs is an experimental new programming model for privacy-preserving computation and AI. It enables rapid compositional development and probable privacy via data flow analysis. Before the 2023 layoffs, the project internally debuted as a core AI-safety system for ambient computing.

• Extended Arcs data flow analysis system to verify MediaPipe graphs. This was an important step towards provably private machine learning applications.
• Created a system for automatic claim deduction in a SQL-like subset of the Arcs language. Claim deduction is a core routine to prove that a program adheres to a privacy policy (Kotlin, Visitor Pattern).
• Added features to the project’s domain-specific language; specifically, type variables, maximum-valued types, and reflection. Together, this helped ship a compile-time privacy checking system into production on Android (Typescript, Kotlin, Data Flow Analysis).
• Created a key compiler component to facilitate allocation of modular programs across distributed computing environments, bridging our web technology codebase to Android. (Typescript, Bazel, Kotlin).
• Extended Google’s build system to support Kotlin to Wasm compilation; created code generators to simplify programming with Kotlin on the web (Kotlin, Wasm, Bazel).
• Developed prototypes to discover machine-learning capabilities within the Arcs programming model (Typescript, Tensorflow JS).

Pre-Google Experience

Aira helps blind and low-vision users access visual information via remote assistance with smart glasses.

• Built core dialog engine for an Android Voice-UX. This allowed our blind and low-vision users to pair Bluetooth devices, call a remote assistant, and rate call experiences conversationally. (Java 8)
• Created a visual-question-answer research prototype for an NSF grant. The system used real-time object detection and rule-based NLP to investigate assistive user experiences for blind and low vision people. (Tensorflow, OpenCV, YOLO)
• Technical lead for prototype of indoor navigation system using computer vision (OpenSfM, ArcGIS).
• Trained and productionized a mobile USD currency classifier model to help blind users identify paper bills. (Tensorflow, Android, MobileNet, Firebase MLKit)
• Created an image tagging game to label integral internal datasets. (Spring Boot, Vue.js, Typescript)
• Led agile rituals such as daily stand-ups, sprint planning meetings, and retrospectives. Started a machine-learning brownbag lunch series.

Dr. Vecchio's research group focuses on advanced materials discovery and their translation to industrial applications.

• Used neural networks and boosted tree based algorithms to make materials science discoveries (Tensorflow/Keras, ResNet, XGBoost, Sklearn).
• Applied gradient-based techniques to explain classifications of convolutional neural networks (GradCam).
• Used OOP design principles to create a framework to track hyperparameters & ML pipelines in git.
• Taught nanoengineering graduate students machine learning, focusing on neural networks and tree algorithms.
• The papers produced as a result of my contributions would go on to be published in Science, Nature, and respected journals within the materials science world.

ProSeries is intuitive tax-preparation software for tax professionals.

• Reduced rate of application crashes in ProSeries by 50% from prior year via data analysis. (Python, Pandas)
• Won business unit hackathon with novel idea for a telemetry system, presented to senior leadership.
• Contributed to annual releases of tax preparation software for legacy Windows application. (C++, C#)

Select Publications

Please visit my Google Scholar page to see my latest publications.

• Learning skillful medium-range global weather forecasting. Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, Alexander Merose, Stephan Hoyer, George Holland, Oriol Vinyals, Jacklynn Stott, Alexander Pritzel, Shakir Mohamed, Peter Battaglia. Science 382 (6677), 1416-1421
• Discovery of high-entropy ceramics via machine learning. Kevin Kaufmann, Daniel Maryanovsky, William M Mellor, Chaoyi Zhu, Alexander S Rosengarten, Tyler J Harrington, Corey Oses, Cormac Toher, Stefano Curtarolo, Kenneth S Vecchio. Npj Computational Materials 6 (1), 42
• Crystal symmetry determination in electron diffraction using machine learning. Kevin Kaufmann, Chaoyi Zhu°, Alexander S Rosengarten°, Daniel Maryanovsky, Tyler J Harrington, Eduardo Marin, Kenneth S Vecchio. Science 367 (6477), 564-568
• WeatherBench 2: A benchmark for the next generation of data‐driven global weather models. Stephan Rasp, Stephan Hoyer, Alexander Merose, Ian Langmore, Peter Battaglia, Tyler Russell, Alvaro Sanchez‐Gonzalez, Vivian Yang, Rob Carver, Shreya Agrawal, Matthew Chantry, Zied Ben Bouallegue, Peter Dueben, Carla Bromberg, Jared Sisk, Luke Barrington, Aaron Bell, Fei Sha. Journal of Advances in Modeling Earth Systems 16 (6), e2023MS004019
• Deep learning for day forecasts from sparse observations. Marcin Andrychowicz, Lasse Espeholt, Di Li, Samier Merchant, Alexander Merose, Fred Zyda, Shreya Agrawal, Nal Kalchbrenner. arXiv preprint arXiv:2306.06079
• Deep neural network enabled space group identification in EBSD. K Kaufmann, C Zhu, AS Rosengarten, KS Vecchio. Microscopy and Microanalysis 26 (3), 447-457

°Equal contribution

Technical Skills

Education

B.S. Computer Science, University of California, San Diego, 2011-16
B.S. Cognitive Science (Computation), University of California, San Diego, 2011-16
President, Cognitive Science Student Association
Co-Founder, Data Science Student Society