← craft

ml

embeddings, graphs, and models at scale

at scale the model is rarely the hard part. the hard part is what you're allowed to call a positive pair, and what the graph does to your assumptions once it's big enough that you can't look at it.

Systems

   profiles ──► encoder ──► contrastive space
                              │
                              ▼
                   ┌────────────────────┐
                   │ graph propagation   │  500M nodes
                   │ k-hop, bounded      │
                   └─────────┬──────────┘
                             ▼
                    retrieval / match
                    eval: offline ≠ online

propagation is where the wins and the failure modes both live. an unbounded hop count makes the offline metric look better and the product worse, which is a lesson you only learn from the online number.

Open

how much of what looks like representation quality is actually eval design.