Blog

FORGE: a shared ML platform from cloud to edge

How one AWS-native platform runs a scheduled ML pipeline, a memory-sensitive API, an edge workload and a conventional service without forcing them all into Kubernetes: six shared contracts, three planes that change at different speeds, and the failures that shaped them.

Treverse Recsys: temporal recommendations from learning to serving

How Haggle recommends items that may leave the auction before their history is useful: a temporal graph retriever, offline ranking, immutable releases, and a controlled experiment that raised sales per exposed buyer by 12.62%.

This site is a bucket and a CDN

Plain HTML in S3, CloudFront in front, one function to make the URLs clean. The whole setup for engineering.haggle.com.

Hello from Haggle Engineering

Why a small team running a live auction platform decided it was worth keeping a public notebook.