The Archive / Research & Exploration
Some questions are worth exploring beyond the boundaries of what we already know.
R.01
Artificial Intelligence & Machine Learning
exploringConcepts I'm studying
What I'm exploring
How models actually learn — the mechanics under the demos: optimization, backpropagation, attention, and what evaluation really means when a metric can be gamed by accident.
Why it interests me
A model that scores well is a claim; a training pipeline you can replay and audit is evidence. Machine learning sits exactly where understanding systems and distrusting demos meet.
Concepts I'm studying
Gradient descent and its pathologies; backpropagation as bookkeeping; transformers and attention; evaluation design — what a metric rewards, it gets more of.
R.02
Quantum Computing
exploringConcepts I'm studying
What I'm exploring
A different model of computation itself — information held in superposition, transformed by gates, limited by decoherence. Not the press-release version; the version where you can reason about why Shor's algorithm is fast and why a fault-tolerant machine is still hard.
Why it interests me
Understanding the foundations is the whole game here — there is no improvising around the linear algebra. And there is a second stake: the Quantum Club (co-founded 26 October 2026) makes this exploration pedagogical — learning it well enough to hand it to the next person whole.
Concepts I'm studying
Qubits and superposition; entanglement as a resource; quantum gates and circuits; decoherence and error correction; the shape of quantum algorithms versus classical ones.
R.03
Computational Research & Reproducibility
activeConcepts I'm studying
What I'm exploring
How to make computational experiments trustworthy by default — provenance that captures the full context of a run, caching that makes verification cheaper than re-running, and checks that can tell when a "reproduction" isn't one.
Why it interests me
Reproducibility is usually framed as a virtue; it is actually an engineering problem with an engineering answer. That reframing became NeuroWeave (started May 2026): a graph kernel, run records with code hashes, seeds and environments, an artifact store, run comparison and reproduction checks.
Concepts I'm studying
Provenance — what metadata makes a run replayable; content-addressed caching — what deserves to invalidate a cache; run records as queryable history; reproduction checks — what "the same result" even means.
Code
NeuroWeave is the running experiment: record a run, perturb a parameter, verify the cache fires exactly where it should — and nowhere else.