about me

hi! i'm leon, i'm a research engineer living in stockholm, lived here for almost 2 years now with my fiance(!). big fan of ml research, and just understanding how things work. otherwise i enjoy eating food, making food, traveling and working out.

i've got a background in software engineering, with a master's in compsci. but today i'm focused on language models, reinforcement learning, mech interp, and reasoning. although my experience is limited i'm very keen on health/medicine as a field of application for the above. if you're curious about my current fixations, skim the blog.

i like open source

these are my favorite, some of which i contribute to.

• rl with trl (core maintainer)
• inference engine vllm
• finetuning with unsloth
• ggerganov's llama.cpp (#4484)
• mlx is an array framework for apple silicon (#202 #237 #456 #19 #276)
• tensor puzzles and cuda puzzles from sasha rush
• community driven ohmyzsh
• jax: the lovechild of autograd and xla

i like making things

healthy mix of research, experiments, and scratch projects.

• paper implementations in omni
• this website, including my blog
• mindex, a local semantic search engine for your mind index
• alphax, scalable + fast alphazero in jax. facilitates spmd, simd, jit-composition etc
• exploring limitations of llm-as-a-judge
• llmcontext (fork of original), a context window pressure tester of open llms
• a framework for super-resolution of clinical 4d flow mri. paper.
• alphafour , minimal alphazero for connect four in pure pytorch

i like reading papers

some favorites of mine.

• Physics of Language Models
• On the Latent Language of Multilingual Transformers
• Transformer Circuits
• GLM 4.5
• AReaL
• Part I: Tricks or Traps? A Deep Dive into RL for LLM Reasoning
• Command-A
• DeepSeek-V2
• MuZero
• MHA, MQA, GQA, MLA (by Jianlin Su)
• Activation-aware weight quantization

ideas

things / projects i'm thinking about lately.

• visual reasoning and visual spatial reasoning. non-text cot?
• proper long-context reasoning benchmark. forcing retrieval + reasoning across the context
• async RL framework optimized for single node training