Projects

Applied ML, systems tooling, and game engines.

A more complete snapshot of the work I want people to inspect first: shipped software, research systems, and demos that expose the engineering behind them.

Model improvement platform

Crucible

Platform for uploading private data, building evals, fine-tuning open-source models, comparing candidates, and promoting winners with lineage and reproducibility.

  • Runs locally, on SSH GPU boxes, or on Slurm clusters while tracking run lifecycle and artifacts.
  • Supports model registry workflows, eval coverage checks, failed-run surfacing, and promotion blockers.
  • Includes an agentic Studio operator for inspecting data, drafting plans, launching approved work, and summarizing results.
Live app

Stack: PyTorch · React · TypeScript · Tauri · MCP

Hexagonal chess engine

Chexagon

A 91-square hexagonal chess engine in Go with a playable WebAssembly demo and a search stack tuned for fast tactical play.

  • 500K-900K nodes per second with alpha-beta, PVS, null-move pruning, LMR, and quiescence search.
  • Transposition tables and engine work delivered a 200x speedup over naive search.
  • PyTorch NNUE pipeline with HalfKP features, self-play generation, and quantized int16 inference.
Play demo Code

Stack: Go · WASM · PyTorch · NNUE · Search

ML research

MusicGen steering

Built a MagnaTagATune-based contrasting concept pair dataset generator and trained learnable steering tensors for MusicGen.

  • Generated 3,000+ controlled contrast pairs with audio transforms and layerwise embeddings.
  • Beat DiffMean baselines by at least 50% in the evaluated setup.
  • Validated with CLAP score, Frechet Audio Distance, perplexity scores, and PCA plots.

Stack: PyTorch · MusicGen · Audio ML · Evaluation

Research in progress

Hyperbolic-JEPA

Research project exploring I-JEPA-style self-supervised vision learning with Poincare-ball representation geometry.

  • Replaces the Euclidean prediction objective with a hyperbolic distance objective.
  • Keeps the ViT context encoder, EMA target encoder, predictor, and masking recipe close to I-JEPA.
  • Currently evaluating linear-probe behavior on CIFAR-scale experiments; repo and paper links will be added after submission review.

Stack: Self-supervised learning · Vision transformers · Hyperbolic geometry · PyTorch

Systems project

C web server

A small web server written in C. This site is currently served by nginx, but the project remains a useful systems artifact.

Code

Stack: C · HTTP · Sockets

Live SSH demo

Termio

Multiplayer terminal snake arena hosted over SSH, written in Go with lobbies, bots, leaderboards, rate limits, and persistent stats.

  • Players connect directly from a terminal and join shared real-time lobbies.
  • Game logic is separated from the SSH server and Bubble Tea terminal renderer.
  • Bots use in-memory tabular reinforcement learning with Monte Carlo tree search planning.
ssh termio.aipinto.com

Stack: Go · SSH · Bubble Tea · Game servers