Oliver Perrin · lead ml engineer

Production ML, with the research discipline to back it.

Boone, NC · B.S. CS, App State 2026 · oliver.t.perrin@gmail.com

Machine Learning Engineer & Researcher with a dual focus on empirical research and production deployment. I'm the sole MLE at a stealth, pre-launch startup — I joined as a senior-year intern after my capstone became the initial production system, and I've since stepped into the role full-time, owning all ML architecture, research, and production decisions in direct collaboration with the founder.

On the research side I'm finishing an EMNLP 2026–targeted preprint on multi-task learning, built on top of a 272M-parameter encoder-decoder transformer I implemented from scratch. I also build LiminalML — a structured path to ML/SWE interviews, grounded in the candidate's own work.

Now

Current focus

Working Lead ML Engineer at a stealth, pre-launch startup — sole MLE on a 10-person team (founder + 8 SWEs). SBERT-based attribute/unit harmonization, schema-agnostic ingestion, and a real-time ML operations dashboard now feeding the database, browser extension, mobile app, and web platform.
Building LiminalML — a structured, non-skippable path to ML/SWE interviews across 157+ topics, grounded in the candidate's own resume.
Researching LexiMind — 272M-param encoder-decoder transformer from scratch, MTL across summarization / topic classification / emotion detection. arXiv preprint, EMNLP 2026 target.
Reading Stanford CS336 (Language Modeling from Scratch) and CS224R (Deep RL) — full lectures, notes, exams.
Highlights

Selected work

Picked for depth — full project list and experience are a click away.

Research

LexiMind

272M-parameter transformer from scratch, MTL with learned attention pooling and temperature-based task sampling. Emotion F1 0.199 → 0.352 (+77%).

Read more → Product

LiminalML

A structured, resume-grounded path through 157+ ML/SWE interview topics in a 6-stage flow from intuition to implementation.

liminalml.com ↗ Production

Stealth startup (core MLE)

Shipped SBERT-based attribute & unit harmonization over 125K+ attributes, schema-agnostic ingestion, and a FastAPI+WebSocket ML ops dashboard.

See experience →