AI Researcher

Sasha Rush Portfolio

ML Researcher, Cursor

ML researcher at Cursor in NYC, working on post-training for coding, after a decade (2016-2026) as a professor at Harvard and then Cornell and a 2019-2024 research role at Hugging Face; created "The Annotated Transformer," a line-by-line PyTorch walkthrough of the original Transformer paper (7.5k GitHub stars).

NLPTransformersLanguage Model TrainingReasoning

Location

NYC

Companies

CursorCornell UniversityHugging FaceHarvard University

Credentials

  • Ph.D., Computer Science, Massachusetts Institute of Technology
  • B.A., Computer Science, Harvard University

Skills

NLPTransformersLanguage Model TrainingReasoning

Notable

  • NSF CAREER Award and Sloan Fellowship
  • Helped start the Conference on Language Modeling (COLM) in 2024
  • Co-author of the Cursor Composer 2 Technical Report (2026)
  • GPU-Puzzles teaching repo: 12.5k GitHub stars

What makes it work

A breakdown of the choices that make this portfolio stand out.

1

A paper re-implemented and explained line by line

The Annotated Transformer pairs the original paper's text with runnable code for every component, turning a dense academic paper into something a reader can execute and modify — a format he reused for Annotated S4 and that, per his site, prompted many follow-ups for newer models.

2

A publication record that tracks each shift in the field

Papers spanning OpenNMT (2017) through masked diffusion language models (2024) and post-training for coding agents (2026) show someone who moved with the field's actual paradigm shifts rather than specializing in one and stopping.

3

Academic and industry research held side by side

Five years of overlapping university and Hugging Face appointments (2019-2024), followed by a full move to Cursor, show a research practice that stayed connected to production systems rather than purely theoretical work.

What AI Researchers can take from this

Specific, actionable tips to apply to your own portfolio — no generic advice.

  • If a foundational paper in your field is hard to parse, re-implement it with inline explanation — it can become more widely used than most original commentary.

  • Let your publication list show real range across the field's shifts over time, not one narrow specialty frozen at one moment.

  • Keep at least one foot in applied or industry work alongside academic research — it keeps theoretical work grounded in what production systems actually need.

Projects

The Projects page lists puzzles, a build-your-own PyTorch, and annotated papers, so a visitor can learn most of modern deep learning just from the side projects Sasha Rush kept alongside the research.

Open SourceadvancedSingle long-form project

The Annotated Transformer

PythonPyTorchTransformersLiterate Programming

The 'Attention Is All You Need' paper aligned line by line with a runnable PyTorch implementation, first published in 2018 on the Harvard NLP blog.

Turns a dense paper into code a reader can run and modify, which is proof of deep understanding.

  • Paper text and code interleaved

  • Spawned his later Annotated S4

  • Listed as a paper and as a project

Open SourceintermediatePuzzle set

GPU-Puzzles

PythonCUDANumba

A series of puzzles for learning GPU programming for deep learning, from tensors to kernels.

Teaching material that working engineers actually use to learn GPU programming.

  • Part of a puzzle series with LLM Training Puzzles

  • Exercise format rather than lectures

  • Runs in a notebook

Open SourceintermediateSemester-long course library

MiniTorch

PythonAutodiffTensors

A DIY teaching library that walks students through building a tensor and autodifferentiation library from scratch.

Shows the ability to design a course around building the tools, not only using them.

  • Build-it-yourself PyTorch

  • Course-scale curriculum

  • Part of a long list of open-source teaching tools

Open SourceadvancedMulti-year toolkit

OpenNMT

LuaPyTorchTensorFlowNeural Machine Translation

An open-source neural machine translation system, first built in Lua with Systran and later ported to PyTorch and TensorFlow.

An early research toolkit that outgrew its original lab.

  • Co-developed with industry (Systran)

  • Ported across three frameworks

  • Handed off to outside maintainers

Ready to build your portfolio?

Follow the AI Researcher roadmap — skills, projects, and timeline to get hired.

AI Researcher Roadmap