ML systems and verifiable fleet engineering — inference kernels, model-runtime infrastructure, signed receipts, rigorous evaluation.
sunlitmoon.pages.dev · Verify a receipt · Dispatch 001
Hiring thesis: verified kernel optimization + inference systems — not repository count.
Tools: Python / PyTorch / Triton / CUDA where the problem needs them; Zig for CPU references, harnesses, and binary formats.
Not claiming: GPU production serving, distributed training at scale, Research Scientist track, or “Pure-Zig everything.”
| Repo | Problem | What is proven | Limitation |
|---|---|---|---|
| tokenizers-zig | HF-compatible BPE / WordPiece / Unigram in Zig | 191 tests total (162 pass in a clean clone; 29 real-model tests skip — HF tokenizer.json fixtures not distributed) + property fuzz |
AGPL; WordPiece speedup is synthetic-fixture — not a general HF replacement |
| inference | End-to-end TinyLlama-1.1B CPU inference | Forward path + BF16 kernels + HTTP surface; zig build test |
CPU-only; single-request; no continuous batching / CUDA |
| faiss-zig | ANN: Flat / HNSW / IVFFlat / IVFPQ | Multi-index families + SIMD batch search | Small-N benches; not a FAISS substitute |
| safetensors-zig | Safetensors reader | Structural scan + dtype coverage on TinyLlama fixture | Read path; not a full HF ecosystem port |
| sme-zig | Structure-Mapping Engine reproduction | Canonical analogies + falsification notes | Classical algorithm — research value needs a modern LM experiment |
Aerospace / naval workbench (strip, yard, sovereign-experience) is a separate product narrative — not the frontier-lab pin set.
Verified Kernel Optimization Environment — agents propose kernels; graders enforce compile, functional/numerical correctness, and performance under shape distributions. Report first; more libraries later.
Plan: FRONTIER_LAB_PLAN.md · Claims ledger: EVIDENCE_LEDGER.md
Proof language: unit-tested unless a numbered bench names commit, machine, baseline, and script. OQ / production / SOTA: false.