Portrait of Ethan Bennett

Ethan Bennett

Flowery Branch, Georgia, United States · Remote

me@ethanbennett.dev

linkedin.com/in/ethan-bennett-dev

github.com/EthanBnntt

Download Resume

I'm a machine learning research engineer with a BS in Computer Science from the University of North Georgia. I run controlled experiments on zeroth-order optimization, quantized training, and distributed training systems, and I co-authored peer-reviewed research on microRNA target-site detection at UNG, where I later helped establish the AI lab on the Dahlonega campus. By day I build production software at ClassLink, where I helped ship the OneData product; previously I engineered software at Imagine Learning. My open-source work centers on JAX optimization and training tooling (zerograd), with side projects in Go and TypeScript.

Experience

ClassLink

Software Developer IC1

Clifton, New Jersey, United States · Remote

  • Shipped ClassLink OneData to production, a platform that automates retrieval, unification, and secure transfer of K-12 assessment data enriched with roster and attendance context; built the data-ingestion pipelines and backend APIs behind it (Angular, Python, PostgreSQL, AWS).
  • Helped drive AI-assisted modernization of the OneData codebase with LLM coding agents, improving maintainability and accelerating feature delivery.
  • Contributed 13 OneData source integrations — including ACT, SAT, i-Ready, NWEA MAP Growth, Renaissance Star, Aeries, Infinite Campus, and Roster Server OneRoster — and diagnosed and fixed performance bottlenecks in the ingestion pipelines.

University of North Georgia - Mike Cottrell College of Business

Machine Learning Research Assistant, Bioinformatics

Dahlonega, Georgia, United States · Part-time from February 2025

  • Co-developed MINN (Multi-Input Neural Network), which encodes microRNA–target-site duplex structure, substructures, minimum free energy, and base-pairing probabilities as image inputs processed by parallel CNN branches (Python, TensorFlow/Keras).
  • Built the dynamic-programming duplex structure prediction pipeline that supplies three of MINN's four input channels, modeling microRNA-specific binding constraints instead of relying on raw nucleotide sequences alone, plus the training and benchmarking infrastructure used to evaluate MINN against its baselines.
  • Co-authored "A Multi-Input Neural Network Model for Accurate MicroRNA Target Site Detection" (Non-Coding RNA, March 2025); on an experimentally validated test set, MINN reached AUPRC 0.937, precision 0.873, and recall 0.870, outperforming TargetScan, RNAhybrid, miRanda, RNA22, TargetNet, Mimosa, and TEC-miTarget.
  • Continued contributing after publication, advising Mohammad Mohebbi as he established an AI lab at UNG's Dahlonega campus.

Imagine Learning

Associate Software Engineer

Tempe, Arizona, United States · Remote

  • Shipped features for Imagine Learning Classroom (ILC), the company's digital classroom product for K-12 instruction (Python, Vue).
  • Built and maintained internal APIs and developer tooling, including the employee API and related internal platform services (C#, .NET).
  • Contributed to DevOps workflows supporting internal services, deployment, and platform reliability (Concourse CI).
  • Promoted from summer intern to associate engineer after completing the 2023 internship program.

Summer Software Engineer

Tempe, Arizona, United States · Remote

  • Applied data-science methods to analyze student grade and performance data, exploring learner outcomes and trends across the platform (Python, Pandas).
  • Worked with production datasets to build a clearer understanding of how students engage with and progress through instructional content.

Publications

More on ResearchGate

Projects

zerograd

JAX/Optax library for zeroth-order optimization: ES pseudo-gradients replayed from low-rank factor perturbations, never materialized densely, with distributed multi-device and seed-derived fault-tolerant cluster modes (params never transferred). Found that 4-bit activations improve ES accuracy +15.7pp without a straight-through estimator while hurting AdamW (controlled ViT/CIFAR-10 study), and wrote a fused int8 linear+LUT Pallas kernel bit-identical to the unfused forward, up to 1.7× XLA at K≤256 on H100.

PythonJAXOptaxPallas

knowledge-cat

CLI and MCP server for reading, searching, and validating Open Knowledge Format bundles so LLM agents can work with structured documentation.

GoMCP

tinygrad-gmlp

gMLP sequence model from Liu et al. (2021) implemented in tinygrad and published as the pip package gmlp_tinygrad.

Pythontinygrad

tinygrad-vit

Minimal Vision Transformer for image classification implemented in tinygrad with a training script and MIT-licensed release.

Pythontinygrad

Education

Key Skills

ML & data

JAXFlax (NNX)OptaxPyTorchTensorFlow/Kerastinygrad

GPU & research methods

Pallas/Triton kernelsint8/4-bit quantizationGPU benchmarking (H100)Ablation study designDistributed training

Languages

PythonTypeScriptJavaScriptGoC#RustSQL

Web frameworks

AngularReactVueHonoFastAPI.NET

Platforms & tools

Cloudflare WorkersAWSDockerPostgreSQLSQLiteRedisNode.jsConcourse CIGitLinux

AI & LLM tooling

MCP servers (Go)Cloudflare AgentsAgentic coding (Pi)AI-assisted development

Domains

Full-Stack Web DevelopmentMachine LearningDistributed TrainingP2P NetworkingBioinformatics

Honors & Awards

Siler Scholars · University of North Georgia

Awarded for achieving the highest GPA among UNG students with 50+ completed semester hours.

2nd Place Team · 2024 UNG Mathematics Tournament

Competed on Dahlonega Team 2, placing 2nd in the team division and contributing to UNG's 2nd-place institutional finish.

President's Honor Roll & List · University of North Georgia

Maintained a 4.0 GPA across four consecutive terms (12+ credit hours/semester), earning President's Honor Roll for undergraduate studies and President's List during dual enrollment. Semester records: spring 2024, fall 2023, spring 2023, fall 2022.