Akshay Agrawal

Akshay Agrawal founded marimo to connect interactive data and AI experimentation with reusable Python software. Its central design represents notebook cells as a dependency graph, allowing the same program to run as a reactive notebook, a script, or an interactive web app. CoreWeave acquired marimo in 2025, and Agrawal continues building it there.[2][5][4]

Agrawal traces marimo's origins to his vector embedding research in Jupyter notebooks. He valued seeing data while working on it, but found hidden state, difficult code reuse, and limited interactivity frustrating. One concrete wish was to select points in a scatter plot and automatically receive the selection as a Python dataframe. He teamed up with Myles, and they launched marimo publicly in January 2024.[4][5]

His preference was to make reproducibility part of the programming environment rather than depend entirely on user discipline. In his August 2025 explanation of marimo's dataflow design, he wrote: “At least for the kind of work that I do, I’d prefer a tool that was reproducible by default.”[5]

Marimo models a notebook as a directed acyclic graph of cells: an edge connects a cell defining a variable to another cell that reads it. Running a cell reruns its dependent cells; deleting or modifying one removes its definitions from memory and reruns descendants. A lazy mode instead marks descendants stale. Scripts execute cells in dependency order, while web-app interactions trigger dependent cells through the same graph.[5]

Agrawal explains that the team chose static analysis of variable definitions and references because it makes dependencies understandable and permits correct construction within that model. The choice has a consequential limit: variable mutations, such as list appends, are not tracked. Marimo also prohibits dependency cycles and variable redefinitions across cells. These constraints define its execution model; they do not guarantee the correctness of arbitrary Python code or external effects.[5]

In a March 2025 article coauthored with Myles and Dylan Madisetti, Agrawal explained another decision: store notebooks as Python programs, wrapping cells in functions so importing a notebook does not execute every cell. This supports readable Git changes, script execution, and testing with pytest. The authors acknowledged the tradeoff that visual outputs are stored separately, using caches and snapshots. They described reusable top-level functions and classes as forthcoming in version 0.12.0; Agrawal's later August article demonstrated function reuse.[6][5]

In his October 30, 2025 acquisition announcement, Agrawal framed joining CoreWeave as a way to continue developing marimo with greater resources, expand the cloud-hosted notebook service molab, and offer GPU-accelerated compute. These were his stated plans for the acquisition. He also made an explicit commitment: “The marimo notebook will always be free, open-source, and permissively-licensed.”[4]

Before founding marimo, Agrawal built computer systems at Google and researched convex optimization at Stanford, where he earned a PhD advised by Stephen Boyd. His earlier work supplies technical context for the machine-learning systems and embedding experiments that preceded marimo.[2][3][9][7]

Agrawal coauthored TensorFlow Eager, submitted in February 2019 and published in the 2019 SysML proceedings. The system combines an imperative interface that executes operations immediately with a JIT tracer that converts TensorFlow-based Python functions into executable dataflow graphs. His collaborators were Akshay Naresh Modi, Alexandre Passos, Allen Lavoie, Ashish Agarwal, Asim Shankar, Igor Ganichev, Josh Levenberg, Mingsheng Hong, Rajat Monga, and Shanqing Cai.[9]

Agrawal is a developer and maintainer of CVXPY, a Python modeling language for convex optimization. He also coauthored Differentiable Convex Optimization Layers with Brandon Amos, Shane Barratt, Stephen Boyd, Steven Diamond, and J. Zico Kolter; the paper lists its authors alphabetically.[3][8]

With A. Ali and Stephen Boyd, Agrawal coauthored the 2021 monograph Minimum-Distortion Embedding. It formulates embeddings by assigning vectors to items and minimizing total pairwise distortion subject to constraints: similar items should be close, and dissimilar items should remain separated. Agrawal developed the accompanying PyTorch library, PyMDE, which approximately solves these problems and lets users experiment with supplied or custom distortion functions and constraints.[7][3]