Oliver
Perrin
I’m the founding machine learning engineer at a stealth startup.
My research interests are reinforcement learning and post-training. I also build LiminalML and write about machine learning.
Read my writingCurrent work
Machine Learning Engineer
Founding team · Stealth startup · May 2026–present
Sole ML engineer on a four-person team. I own the ML stack end to end: model training, evaluation, inference, and integration into the product.
The product is pre-launch, so the details are private.
ExperienceLexiMind
Research · Open source · Nov 2025–present
Does model merging (task arithmetic, TIES) match multi-task LoRA fine-tuning once you count the compute spent training each specialist? A budget-matched study on three text-classification tasks, using a transformer I implemented in PyTorch and initialized from FLAN-T5.
Baselines are done; the main runs are next.
LiminalML
Independent product · Apr 2026–present
Bridging the gap between knowing machine learning and understanding it.
LoRA update space
Change the adapter rank to see its outputs expand from a line to a plane to a volume.
A three-output toy model with fixed synthetic inputs. It shows the adapter correction, not trained model activations.
Open in StudioPreparing the interactive visual…
Writing
Implementation notes, explanations, and things I learned the hard way.