Research & systems
I study language-model behavior, training signals, and verification. These projects include methods, results, and limitations, with source material where available.
Upbringing
Language-model post-training experiments using SFT, DPO, and verifiable rewards.
BusyCoq contributions
Contributions to formally verified Busy Beaver proofs, checked in Coq and merged upstream.
Curiosity & satisfaction in LLMs
Activation-steering experiments testing whether a candidate direction changes information-seeking behavior.
More projects
rag-eval-kit
A published Python framework for evaluating RAG pipelines across answer quality, retrieval, latency, and cost.
NeuroTechML
Team research on phoneme classification from surface EMG, including signal processing, LSTM training, and live inference.
Custom LLM Harness
A custom agent harness for managing inference, running tools, recording trials, and comparing results.
llm-scry
A prototype for inspecting token probabilities and model activations through a local TransformerLens API.
Research notes
A closer look at two experiments.
What the checker actually checks
A reward curve, a learning rubric and a formal proof each have a boundary. The useful work starts by making it visible.
5 min readWhen steering changes nothing
A completed model-internals experiment, a verified perturbation and a null behavioral result—and what each does and does not establish.