Engineering
Technical writing, not marketing. Each post is scoped to a concrete claim, a reproducible experiment, or a decision we made that we would have wanted to read before making it.
9 min read
Quantized Tensor Train decomposition is a mathematical idea older than modern vector search. Here is what it compresses, what it doesn't, and why that matters when 100 million embeddings have to fit on a single GPU node.
11 min read
Exact recall at 100M vectors on a single GPU node is the headline, but ANN is the right tool for many real workloads. Here's how to decide which regime you're in, measured against published HNSW numbers.
7 min read
A signature on a PDF is marketing. A signature on a manifest that hashes every artifact is a receipt. This post walks through the schema we use, why every artifact is signed with ML-DSA-65 over an SHA-256 manifest, and what a reader can independently verify.