data science & applied statistics.

c-1n / learning to walk

six-legged locomotion learned with PPO.

walk_fast_500 reaches 0.888 m/s in the recorded mean-policy run, with 0 falls across 24 fixed flat-ground evaluations. PPO learns joint corrections and gait cadence. compare the preserved policies in the film.

scopefoot slip and contact fragmentation remain. · terrain, push recovery, and energy efficiency are unproven.

lm lab

how can a language model be built and checked from its parts? an ai-assisted learning lab in pytorch.

byte-pair tokenization, bigram, attention, and transformer components in pytorch. the examples below are preserved notebook outputs from revision 19222f1.

tokenizer notebook ↗

unicode round-trip assertions; a 512-token vocabulary encodes the first 100,000 characters of tiny shakespeare into 47,586 tokens.

bigram notebook ↗

1,000 adamw steps. one validation batch records cross-entropy 4.376, against 6.238 for a uniform distribution over the vocabulary.

attention notebook ↗

two heads; input and output [32, 8, 32], attention weights [32, 2, 8, 8]. causal masking, pre-normalization, and residual connections.

scopein these recorded examples, attention is checked at the block level and the trained model is the bigram baseline. newer source work includes a transformer comparison harness.

log pose

what can dated source inventories reveal about a software market?

python ingestion preserves source hashes, capture dates, and arrival times in postgresql; cutoff queries select eligible captures. a read-only graph API connects pinned inventory co-listings to their exact retained records, with dated search and source audits in the browser.

scopeco-listings are same-category, same-year observations, not verified company relationships or market share. a shared-exposure hypothesis remains untested.

indigo circuit

what can tournament results tell us about player and deck performance?

tournament results feed a nightly duckdb/dbt pipeline with glicko-2 ratings. the dashboard validates shadow-database rebuilds before promotion, with a direct-write fallback. its experimental deck estimates use bayesian shrinkage and approximate 90% model intervals that exclude baseline uncertainty and dependence between cards; associations are not causal card effects.

fourier drawing

interactive fourier reconstruction of drawn curves.

drawn 2d curves are reconstructed with rotating vectors. the interface exposes the periodic seam and normalized rms reconstruction error.

scopeerror is measured against the input curve.

voidpulse / jellyfish

audio-reactive webgl jellyfish.

procedural mantle, filaments, oral arms, and accretion disk. local files or captured tab audio drive the full application, with separate playback and visual-intensity controls.

scopethe black-hole lensing is an artistic approximation. this page shows the silent version; the full app responds to audio.

tiramisu

a searchable lyric-sheet reader with focus mode and Genius annotations.

plain-text lyrics sit beside a decorative p5 cherry-blossom canvas, with focus mode and a separate panel for public Genius lyric annotations. reading remains usable without animation.

receipts

campaign funding and recorded decisions, with public sources.

browse 539 congress members and 532 campaign finance summaries. inspect dated funding, matched roll calls, selected financial disclosures, and candidate comparisons with links to the official records.

scopebundled dated snapshots, not live feeds. votes and disclosures have limited stated coverage; missing data stay unknown. these records do not establish motive or influence.