High-performance GPU/CUDA on NVIDIA Blackwell (sm_120), across three fronts:
- Number theory & cryptography — Miller-Rabin primality, prime gap searches, Pollard's Kangaroo (ECDLP on secp256k1), and empirical verification of the Beal conjecture.
- Differentiable 3D geometry — techniques from scroll-unrolling (winding coordinates, invertible-by-construction maps), applied and stress-tested outside their origin to find where they hold and where they break.
- Vesuvius Challenge — data augmentations for the Herculaneum scroll pipeline (May 2026 Progress Prize - https://scrollprize.org/winners).
- mr_blackwell — Native Miller-Rabin CUDA kernel for NVIDIA Blackwell. CGBN replacement using Montgomery CIOS and PTX carry chains.
- PSCKangaroo — GPU-accelerated Pollard's Kangaroo for secp256k1 ECDLP. Fork of RCKangaroo with concurrent mode, crash-safe checkpoints, and compact 16-byte DPs.
- verimath — Deterministic, third-party-verifiable number-theory agent (factorization, primality proofs) on the CROO Agent Protocol. Every result ships with a reproducible SHA-256 attestation.
- beal_bigint — GPU search for Beal conjecture counterexamples using the by-C^z parametrization. 20K+ hits validated, 0 counterexamples found.
- winding-coordinates — Differentiable winding/shifted coordinates from 3D scroll-unrolling, with the technique's applicability boundary measured: single vs. nested winding, helix vs. phyllotaxis. Live in-browser visualizer.
- icicle-blackwell-ntt — Empirical characterization of ICICLE NTT on consumer Blackwell (sm_120); profiling-driven, with a prototype for the digit-reversal bottleneck.
- bitcoin-puzzle-research — Documented dead ends and empirically refuted approaches in Bitcoin puzzle research.
Empirical before theoretical. Negative results published with the same rigor as positive ones.