cv

Stanley Ngugi

I build RL environments for model post training.

I'm Stanley. I'm 20, self taught, and an independent AI researcher working on reinforcement learning and post training. Much of my work so far has focused on mathematical reasoning and coding, particularly building RL environments and designing rewards for these tasks.

CV github rss

Selected writing

An RL Environment Where C Code Has to Be Proved, Not Just Tested RL, formal methods · 20 min
An RL environment where models write C functions and reward comes from proving that the code satisfies a fixed specification, rather than from passing a test suite. The public release includes 64 tasks and an isolated Frama-C judge.
Grading Mathematical Answers Without Answer Keys RL, formal methods · 23 min
A math RL environment for bounded integer problems that does not store expected answers. Models return an integer or a complete finite certificate, and reward comes from checking that submission against the full encoded problem.
Building a Lean-Backed Verifier for Bounded Mathematical Answers formal methods · 16 min
A verification engine for bounded integer problems. It turns a problem specification and candidate answer into executable Lean code, checks the entire finite domain, and gives a definitive result without requiring a proof from the model.
How a Grammar Changes AI-Generated Mathematical Proof Steps formal methods · 26 min
A study of how grammar masking keeps language-model generations inside the shape of Lean tactics, preventing malformed proof steps without confusing valid syntax with proof correctness.
Building a Grammar for AI-Generated Mathematical Proof Steps formal methods · 19 min
A practical study of Lean's tactic language: extracting tactic forms from real proofs, designing a grammar around them, and stress-testing what the grammar accepts and rejects.
Taming Incidental Polysemanticity in Toy Models interpretability · 25 min
How training choices shape feature entanglement in overcomplete toy networks, using sparse autoencoders to compare the representations learned across different configurations and random seeds.

Earlier research

Unlearn, Then Learn: A Circuit-Informed Knowledge-Editing Experiment preprint
An early experiment in changing one factual association in Phi-3-mini: first locate and suppress the original association, then train a replacement. The complete experimental pipeline is public.
Targeted Lexical Injection: Early-Layer LoRA for Swahili–English Alignment preprint
An early experiment using small adapters to strengthen Swahili–English word alignment already present inside Lugha-Llama, evaluated on training pairs and a separate set of unseen words.