Systems engineer building ML infrastructure from first principles. Inference kernels, vector search, tokenization, agentic verification, and quant tooling. Zig/Elixir/OCaml.
✉️ sean@sunlitmoon.online · 🔗 sunlitmoon.online
| Repo | What it is | Evidence |
|---|---|---|
| inference | Pure-Zig LLM serving — paged KV-cache, BF16 kernels, persistent thread pool, safetensors integration. TinyLlama-1.1B end-to-end on CPU. | 77 tests · 6.17× decode speedup at M=8 vs baseline · STATUS.md canonical |
| tokenizers-zig | Pure-Zig HF tokenizers — BPE, WordPiece, Unigram, full pipeline, offsets, tokenizer.json compat. |
189 tests + 600-iter fuzz · ~5.3× faster on WordPiece · parity verified vs 🤗 tokenizers |
| faiss-zig | Pure-Zig ANN — Flat, HNSW, IVFFlat, IVFPQ; SIMD @Vector distance kernels, multi-threaded searchBatch. |
68 tests · 16.94× memory compression on IVFPQ at reference recall · cross-validated vs FAISS C++ |
| safetensors-zig | Pure-Zig safetensors reader — @Vector(32,u8) structural scan, BF16/F32/I8, libc-only. |
21 tests · 241µs parse on 201-tensor TinyLlama fixture · upstream fixture contribution |
| Repo | What it is | Evidence |
|---|---|---|
| evals-agentic-control | Pre-registered adversarial prompt battery for agentic computer-use safety — 40 prompts × 5 categories × 2 models. Two-stage verdict pass: rule-based + LLM-as-judge on unclear cells. 94.4% safe (CONFIRMED ≥ 80%). 4 confirmed full-compliance cells in C5-cron-social-pressure. Judge Type-II rate on full-compliance label documented. | 161 cells · verdicts.jsonl + combined-summary.json · 7 Type-I/II catches documented |
| stax-mast | Single-binary editor kernel — buffer-as-protocol, Janet (Lisp) extensible, in-process capability sandbox, adversarial verification loop. | 64 tests · 65 Janet bindings removed/gated · mutation-tested security check · AGPL |
| Repo | What it is | Evidence |
|---|---|---|
| quant-validation-zig | Bailey–López de Prado bias-defense: Probabilistic & Deflated Sharpe Ratio, Purged + Combinatorial-Purged K-fold, CPCV backtest path distribution (positive_fraction anti-luck check). | 46 tests · erfc to ~3×10⁻⁸ vs scipy · reproducible Nix builds |
| zsym | Rigorous statistical cryptanalysis substrate — Miller-Madow entropy, n-gram LMs, monoalphabetic hill-climber, polyalphabetic (Vigenère) IC period detection + per-position solver, stationary bootstrap CIs. Applied to Voynich EVA. | 37 tests · bit-exact parallel/serial parity · deterministic seeding · Python parity verified |
| zig-h3 | Uber H3 v4 geospatial index — idiomatic wrapper + pure-Zig port of all 70 public functions. | 211 tests · 27k+ cross-validation cases · 94.4% coverage · property/fuzz/mutation |
| rippled-zig | XRPL protocol toolkit — canonical tx encoding, secp256k1/Ed25519 sign+verify, live testnet RPC conformance. | 5-gate gated process (build → serialize → crypto parity → live RPC → security) |
- NixOS/nixpkgs — CrowdSec hermetic-deploy option; ydotool uinput autoload (open)
- HuggingFace/tokenizers — BPE legacy-merge disambiguation fix
- FAISS (facebookresearch) — IVFPQ recall documentation fix
- safetensors — Llama-3.2 shape-fixture generator; header whitespace-tolerance spec
- PyO3 · clap-rs/clap — docs and builder-method contributions
- Evidence vocabulary is fixed.
sketch → compiled → unit-tested → integration-tested → audited → benchmarked. Words like production or verified require the evidence that makes them true. - Pre-register before claiming. Falsifier written before the test runs. Type I (overclaim) and Type II (missed risk) tracked as separate error classes. Inconclusive results reported honestly.
- Tests are the spec. Priors update when proved wrong.