Why we build on the edge: lessons from shipping Huki

Personalization has a speed problem: if ranking a page takes too long, the shopper is gone. That's why Huki runs its decision logic on edge infrastructure, close to every shopper.

What we've learned

• Keep the hot path small: score with compact vectors, not heavy models.
• Store personal state with the person when possible — it's faster and more private.
• Move heavy work (embeddings, catalogue scoring) to background jobs.

The same architecture now powers our client projects, from storefronts to recommendation APIs.