Projects

LSF, Lightweight Study Format & tools

Status: 1.0.0 released

An open file format for storing and sharing coursework, built for self-study. The main idea is to store the past papers, the marking rubric, your attempts and any media in a single format. I cant do it justice in this description, just read the SPEC.md.

Tooling comes next: parsers and libraries, then an editor, marking tool, merger and a full app (Pivot). A project for summer 2027.

• Specification: github.com/Sardiniangale/lsf-spec
• Where the tooling will be: git.dyno.cx/Sardiniangale/Pivot

Duck

Status: In progress, early

A compatibility tool for TUI and basic apps that runs local binaries instead of containers. Long-term light project I want to work on casually

• Git repository: git.dyno.cx/Sardiniangale/Duck

SHAUN

Status: Operational

A rack of multiple low-power mini Optiplexes in a cluster. Used for embarrassingly parallel simulations and long-term computation over weeks.

• Dedicated page: SHAUN project page

Acqua Calda

Status: Rebuild and Overhaul

A planned high-performance DEM framework for heat transfer simulations in Julia.

Current plan to implement by release in two years. I want to achieve the following from the project:

• Mechanical contacts (Hertz-Mindlin, linear spring-dashpot and the like)
• Particle-particle and particle-wall heat conduction
• Frictional heating and energy conservation
• Interface for custom contact models, thermal solvers, and I/O

• Active again now that LSF 1.0 is out
• Git repository for Acqua Calda: github.com/Sardiniangale/AcquaCalda.jl
• Git repository for VeloceRNG (early planning): git.dyno.cx/Sardiniangale/VeloceRNG
• Internal design docs: sardiniangale.github.io/AcquaCalda.jl

Chess bot

Status: Active and minor upgrades

A chess bot I call u_mad due to its extremely aggressive nature, which was intended.

• Design: HalfKAv2 NNUE, 40,960 input features mapping to 256-dim hidden layers, 10M parameters
• Training: 200,000+ filtered Lc0 positions, with self-play reinforcement learning. Reward shaping with time-decay function R = +-(result) × (1 + e^(-λM)) favors fast, decisive wins. Short games weighted up to 2x
• Eval: Dynamic contempt scaling from -150cp in the opening to 0cp in the endgame.
• Search: Alpha-Beta with principal variation search, 1M-entry aging transposition table, late move reductions and iterative deepening with aspiration windows.
• Book: Polyglot opening book with aggressive lines, dropped after 4-6 moves depending on game mode
• Git repository: git.dyno.cx/Sardiniangale/u_mad
• Go against it here: lichess.org
• Account here: lichess.org

Community infrastructure: see the Starfed projects & blog page.

More projects are always in the works. Email me at giacomo@cadaxity.it for any specific info.

Home|Contact