I build the infrastructure that makes language-model systems dependable.

I’m an AI/ML engineer and researcher. In my own words, I build agentic frameworks, cognitive runtimes, and evaluation systems for LLM applications. The work is public: a cognitive runtime published on PyPI that gives agents typed, budgeted memory with supersession and quarantine; a single-file agent framework with 488 offline tests and a fail-closed authorization path; and a chess analysis engine, live on Hugging Face Spaces, whose selected numerical and move-label claims are checked against engine output before the commentary is shown.

Around those systems I’m building an experimental platform for context engineering with controlled baselines and effect sizes, a multi-agent language tutor built around a transparent learner model, a 104-notebook mathematics curriculum for AI, and an interactive algorithms knowledge base that runs Python in the browser.

My research interest is how the structure of a dataset — its balance, diversity, size, augmentation and preprocessing — may shape what convolutional and attention-based networks learn and how well they generalise. I present that work here as a proposed study design, not as completed research. The same habit runs through my engineering: change one thing at a time, measure generalisation rather than fit, and publish limitations with the evidence.