Founder of LuxiEdge
Deterministic computing, measured in public.
I build numerical and AI systems for work where a result should be reproducible—not merely close—when execution conditions change.
LuxiEdge began with a practical obsession: remove unnecessary numerical work, reduce data movement, and make the path through a computation explicit. Determinism emerged from that work and became the organizing principle.
Today I am building deterministic numerical systems across AI inference, quantitative computing, scientific workloads, and GPU primitives. The commercial engines remain private. The measurements, evaluation binaries, hashes, receipts, limitations, and reproducibility artifacts are published where they can be checked.
The standard: state exactly what was tested, preserve the evidence, and publish the loss when the incumbent wins.
| Project | What you can inspect |
|---|---|
| LuxiDemo | The public evidence room for LuxiEdge: evaluation binaries, checksums, receipts, benchmark methodology, raw run artifacts, and scoped claims. |
| Qwen2-7B inference results | Batch-invariant inference on an NVIDIA H100, with matched comparisons against vLLM and SGLang, plus energy and correctness evidence. |
| Art of Fugue | A small cross-platform demonstration of floating-point drift—and the deterministic alternative. Start here for the problem in its simplest form. |
| Deterministic PairHMM | A methods note, DOI, evaluation binary, and recorded CPU/GPU evidence for configuration-stable score reduction. |
| Attention V2 Benchmarks | Frozen H100/H200 artifacts and a public verifier. It includes the cases where the experimental approach wins and where Flash wins decisively. |
- Deterministic LLM inference — the same prompt produces bit-identical tested results across batch sizes and repeated runs, without accepting the usual deterministic-mode performance penalty.
- Deterministic numerical computing — portable execution policies, canonical reductions, exact output receipts, and evidence that travels with the result.
- Quantitative and risk workloads — reproducible books, scenarios, statistics, and risk-state computation.
- GPU efficiency — less data movement, fewer unnecessary trips through memory, fused work, and honest energy measurement.
- Scientific computing — reproducible reductions and numerical behavior in genomics, orbital mechanics, and related workloads.
I also contribute focused fixes upstream when the problem belongs in the public ecosystem.
- rust-bio PR #688 — fixed a PairHMM global-banding origin bug that could score exact matches as impossible; merged September 2026.
- mamba2-context-parallel — public work on context-parallel state handling, numerical symmetry, device selection, and validation.
- ocaml-vcd-parser — a focused hardware-trace parsing project.
These projects belong here on purpose. I build tools for businesses, family, learning, cooking, energy, and whatever problem is close enough to touch.
| Project | Why it exists |
|---|---|
| Rylee’s Lost Creek Ranch | A fourth-grade math adventure built for my family, with ranch jobs, progress tracking, and curriculum-aligned practice. |
| Recipe Tables | Turns recipes into clean tabular cooking plans, supports scaling and substitutions, and exports shareable PDFs. |
| Free EUR Predictor | A free, physics-informed early EUR forecasting tool for unconventional wells. |
| Recipe Cards | A simple public home for useful recipes and the artifacts produced by Recipe Tables. |
Some repositories are open-source projects. Others publish evaluation binaries and evidence while the commercial implementation remains proprietary. Each repository states its own scope, license, hardware, and limitations.
If you want to evaluate LuxiEdge against a real workload, bring the workload and the incumbent. We will define the kill bar before the run.