Systems engineer. Pure-Zig ML infrastructure, inference kernels, and verifiable tooling.

Writing: sunlitmoon.online · Models & datasets: huggingface.co/SMC17 · sean@sunlitmoon.online

• inference — pure-Zig LLM serving: paged attention, BF16 kernels, persistent thread pool, safetensors integration. TinyLlama-1.1B end-to-end. 115 tests.
• tokenizers-zig — BPE, WordPiece, Unigram with tokenizer.json compatibility and offsets. 189 tests + 600-iteration fuzz.
• faiss-zig — Flat/HNSW/IVFFlat/IVFPQ ANN with SIMD kernels. 16.94× memory compression on IVFPQ. 76 tests. No C/C++ dependency.
• safetensors-zig — safetensors reader with @Vector structural scan. 241µs parse on a 201-tensor TinyLlama fixture. 21 tests.
• zkdb — columnar time-series database with q/kdb+ semantics: typed vectors, SIMD aggregation, asof join, q IPC protocol. 24 tests.
• zig-h3 — Uber H3 v4 geospatial index: idiomatic wrapper + pure-Zig port. 211 tests, 27k+ cross-validation cases, 94.4% coverage.

Claims ship at proof level: test counts come from the runner, performance numbers ship with the benchmark that produced them.