Three years of MECE trees and PRDs taught the same skill twice — but product's feedback loop, shipping to millions instead of handing off a deck, is a faster and less forgiving teacher.

A design pattern for AI-generated business reports: separate the data layer, the analysis-standards layer, and the presentation layer so that a metric's definition only has to change in one place, not five.

Why a vector database plus RRF reranking gives you a search engine, not a memory system — and the five-level architecture I designed to give agents time-aware, causally-linked, self-compressing memory instead.

Taobao's bet that an AI search agent can win back the fuzzy, high-consideration queries it's been losing to Xiaohongshu and Douyin — and why its first gray-test results came in well short of target.

How 1688 built an AI 'digital employee' for its long-tail B2B merchants — the memory architecture behind it, the team structure that shipped it, and why an 'invisible' AI experience beat a configurable one for this audience.

Why enterprise agents built by the business itself get judged on outcomes (revenue, GMV, conversion) while agents sold by SaaS vendors get judged on usage — a metrics teardown across Alibaba's AIBD, 1688, customer-service agents at three major platforms, and a half-dozen B2B sales-agent startups.

A plain-language walkthrough of how large language models are actually trained — pretraining, SFT, and RL — paired with a framework for how enterprises of different sizes should choose their own LLM strategy, illustrated with real deployments from Bloomberg, an ed-tech company, and retail chains.

A survey of how 2B agents are landing across e-commerce, advertising, and local services, paired with a deep dive into agent memory — why B2B, enterprise, and consumer users each demand a completely different memory experience.