Physics, math, and AI - built in the open.
Navy RF systems → EPA AI policy → AI/ML since 2019. I build tools and do research at the intersection of AI, scientific computing, and art. Most of it starts as a question I can't stop poking at and turns into code, a paper, or both.
- graph-graph - do simple local graph rules at large scale produce interesting emergent behavior? Sparse simulation out to millions of nodes, dimension analysis, and experiment waves that are pre-registered before they run.
- terravector - terrain patch similarity search: decomposition-based embeddings of elevation data, indexed with HNSW. Grows out of RESIDUALS, my differential feature-detection toolkit for archaeological prospection.
- alcubierre - reframing the warp-drive metric's exotic-matter requirement as a boundary (Casimir-like) effect. Phase-1 linearization verified symbolically and numerically; v0.3.0 archived on Zenodo.
- ESTA - a local LLM wrapper that exposes auditable internal state (token confidence, refusal-direction projection) alongside OpenAI-compatible responses.
- integer-sequence-visualizer - live OEIS visualizer with a null-model layer: is the structure in the sequence, or in the rendering?
- rna-joint-normality - proof of a Bu–Kauers–Zeilberger conjecture: hairpins and basepairs in random RNA secondary structures are asymptotically jointly normal.
- 3body-poisson-algebra - the dimension of the pairwise Poisson algebra of the N-body problem (3, 6, 17, 116…). Dashboard → nbody.briansheppard.com.
Topological data analysis on survey astronomy, physical learning substrates (FPAA / memristive / Ising), retro ASIC characterization, bioelectric memory, generative art, and a lot of developer tooling. Some of it ships; some of it is an honest dead end I left standing on purpose.
Curiosity-driven, adversarial about my own claims, and allergic to copy that runs ahead of the code - I'd rather publish the failure than the press release.
I also run Dark Forest Labs and keep a few tools at bshepp-projects.