Skip to content
View sweeney-007's full-sized avatar
  • San Jose, California
  • LinkedIn in/mlsween
  • Joined Nov 8, 2026

Block or report sweeney-007

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
sweeney-007/README.md

Hi, I'm Mitja Leb 👋

AI engineer who likes understanding models from the inside out: I build transformers, tokenizers and training pipelines from scratch, run controlled experiments to find out what actually matters, and turn LLMs into reliable applications with verifiers, guardrails and evals.

My projects follow the questions behind recent progress in AI: how to pick architectures and hyperparameters on small models that still hold at scale (µP, mixture of experts, scaling laws), how to make model output checkable (verifiers, execution feedback, fact-checking before answering), and how models can learn from their own verified outputs.

  • 🔭 Currently working on: verifier-guided code generation for C (SEGVibe) and scaling experiments (tinyformer)
  • 🌱 Learning: RL from verifiable rewards, LLM evaluation, efficient training and inference
  • 📫 Reach me: LinkedIn

Featured projects

Project What it is Highlights
tinyformer A transformer and byte-level BPE tokenizer from scratch (Python + raw PyTorch), plus µP and mixture-of-experts scaling experiments on cloud GPUs BPE beats character tokens by 31% in bits per character; under µP the best learning rate stays put across an 8× width range; MoE beats dense at every width per FLOP but not per GPU-second.
SEGVibe A code LLM writes C; sanitizers, fuzzing, static analysis and profiling check it; findings go back for repair The checks catch 93% of 95 seeded defects with 0 false positives; 12 defects are caught only by a layer other than the tests. Includes a LangGraph repair agent and retrieval of CWE fix guidance.
filings-analyst LLM agent that answers questions about company annual reports (SEC 10-K) with citations, built with LangGraph Hybrid RAG with corrective retrieval, tool calling, fact-check step, typed error handling. 93% hit@5 retrieval on 300 questions.

Tech I use

Python · PyTorch · C · Hugging Face · vLLM · LangChain · LangGraph · FastAPI · Docker · Modal · Clang sanitizers · Git · Linux

Pinned Loading

  1. tinyformer tinyformer Public

    From-scratch GPT and BPE tokenizer in raw PyTorch, plus µP and mixture-of-experts scaling experiments on cloud GPUs. BPE beats character tokens by 31% in bits/char; µP keeps the best learning rate …

    Python