Resume

Finn Eitreim — Machine Learning Engineer

Work Experience

Zed — Open Source Internship Remote · March 2026 — May 2026

Selected for the inaugural Zed Guild cohort, contributing directly to the Zed editor codebase in Rust.
20+ merged PRs, ~20% of the merged PRs for the whole cohort.
Traveled to RustWeek in Utrecht, Netherlands.

Western Washington University — Machine Learning Research Assistant Bellingham, WA · January 2022 — December 2025

Led Timescale Disentanglement project: temporal analysis and neural video compression.
Took ownership of project direction after senior graduate students departed; contributed to research paper drafts.

Western Washington University — Full-Time Research Intern Bellingham, WA · June 2025 — September 2025

Engineered high-throughput data pipelines for NVIDIA Hopper accelerators with hardware video decoding.
Optimized model training loops and profiled GPU utilization across multi-node configurations.
Built visualization tools for high-dimensional latent spaces using dimensionality reduction.

Western Washington University — Full-Time Research Intern Bellingham, WA · June 2024 — September 2024

Developed and iterated on novel architectures based on VQ-VAE, GANs, and Diffusion models.
Replicated and extended methods from current computer vision literature as evaluation baselines.
Designed latent space metrics and evaluation systems for rigorous, reproducible experiments.

Ferro-Kittens — Blackwell Tile Library — ferro-kittens

ThunderKittens-style tile abstractions in Rust for Blackwell: typed shared/register/TMEM tiles and tcgen05 MMA.
Wrote cluster GEMM, flash-attention, softmax, and layernorm kernels, benchmarked against cuBLASLt on B200.
Built ptxas-based register/spill/occupancy gates that run in CI on every PR without needing a GPU.

Oxide-Train — LLM Training Engine in Rust — oxide-train

Built a from-scratch LLM training engine in pure Rust on NVIDIA B200, with GPU kernels written in Rust via cuda-oxide.
Trains a ~4B-parameter (1B active) MoE model on Wikipedia; implemented data sharding, tensor ops, backprop, and AdamW.
Built a per-kernel CUDA-event profiler with same-container baseline comparisons to catch performance regressions.

Spherical Latent Flow Model — flow-sphere

Designed a two-stage generative model: ViT-based spherical VAE followed by a DiT flow refinement step.
Constrained latents to a hypersphere via RMS normalization; initialized flow from AE output for few-step convergence.
Trained with Muon optimizer on weight matrices and AdamW elsewhere; evaluated with FID on CIFAR-10 and ImageNet.

AI Image Detector — detect-fake-images

Built a paired real-vs-fake dataset: ~41k COCO photos matched with FLUX.2-klein generations from the same captions, jpeg-matched 64×64 crops.
Trained detectors from a 25-parameter logistic regression on hand-crafted features up to a tiny ViT reaching 95.4% accuracy / 0.991 AUROC.
Averaged per-crop scores for free test-time ensembling, roughly halving image-level error without retraining.

3DGSforMLX — Gaussian Splatting on Apple Silicon — 3DGSforMLX

Built a 3D Gaussian Splatting trainer in MLX, with a differentiable rasterizer written from compiled, vmapped MLX ops.
Replaced the MLX compositor with a hand-written Metal kernel: front-to-back alpha compositing with transmittance-based early termination.
Trained with a DSSIM loss and staged spherical harmonics, from point-cloud or random initialization.

CuTe-rs — implementation of the CuTe layout algebra in Rust.
Tome — LLM inference engine with a Rust TUI; a testbed for serving-optimization and quality experiments, wired into OpenCode for real-world use.
value-rl — local reinforcement learning system teaching intellectual curiosity, nonsense detection, and claim scrutiny; optimized for Apple Silicon.
PanGAN — adversarial RL training against the Pangram AI detector.
esp-digits — digit generation with a simple diffusion model, embedded on an ESP32.
safe-kernels — GPU kernels (GEMM, reductions, vector ops) in pure Rust with cuda-oxide; covered in a blog post.
cholesky_factorization — fast, zero-allocation Cholesky decomposition in Julia.
AutoCUDA — LLM-generated C++ torch operators from PyTorch functions, with automatic correctness checking and benchmarking against the baseline.
timescale-diffusion — research code from the Timescale Disentanglement project.
website — this site: a static site generator + web server in one small Rust binary, with build-time LaTeX-to-SVG math.
light-audio-player — lightweight audio player in Rust.

Western Washington University — B.S. Computer Science Bellingham, WA · September 2021 — December 2025

Languages & Frameworks: Python, Rust, PyTorch, JAX, MLX, CUDA, Metal, C/C++, Java, Julia, R

Machine Learning: Computer Vision, Variational Autoencoders, Diffusion Models, Transformers, 3D Gaussian Splatting, High-Speed Data Processing

Development Tools: Git, Linux, High-Performance Computing, Embedded Programming, Embedded Debugging, Vim