I spent fifteen years making systems bigger.

A billing pipeline that went from 50 to 10,000 queries per second. A data platform processing five billion records a day. An engineering org that grew from 20 to 150 people in six months. Then I moved to Canada, started over, and found that the most interesting problems now fit in a single SQLite file.

What didn't change is the itch. Every new field, whether RAG, agent runtimes, the Apple Neural Engine, or edge databases, I learn the same way: build something real with it, find where it breaks, write down what held up. This blog is about what scale taught me. Mostly, it taught me restraint.

Then · 2008–2023 · Beijing Now · 2023– · Waterloo

10,000 QPS ad-billing engine, rebuilt from 50

1 file SQLite in WAL mode, holding an agent's memory
5B records/day on a Hadoop data platform

ms latency multi-hop BFS, in pure Python
150 engineers shipping to 30+ cities

1 engineer working alongside AI agents
1.7 TB CRM migration for H&M China, 40+ tables

0 servers edge functions and local-first tools

What I believe, with receipts

01 Scale is a bill, not a badge.

Receipt Rebuilt a failing ad system from scratch in one month; taking it from 50 to 10,000 QPS taught me exactly what every extra nine costs. (Meili Group, 2013)

Evidence Graph Traversal Without a Graph Database: BFS in Pure Python: where I argue Neo4j is overkill for an agent's memory.

02 The best dependency is the one you didn't add.

Receipt Built ZEWS, an in-house cloud monitoring system at Zynga, instead of waiting for tooling to arrive; it earned a headquarters commendation. (2012)

Evidence Building a Local-First Memory Layer for AI Agents: no server, no vector store, no graph database.

03 Concurrency is a design problem before it is a tooling problem.

Receipt Designed multi-tenant financial settlement for an ERP serving dozens of cities; isolation lived in the schema, not in middleware. (2018–2020)

Evidence Making SQLite Multi-Agent Safe: WAL mode, thread-local connections, and one well-placed RLock.

04 Dirty data compounds like debt.

Receipt Migrated 40+ tables and 1.7 TB of CRM data for H&M China under regulatory compliance; every shortcut would have surfaced in an audit. (2020–2023)

Evidence Building a Deduplication Pipeline for Local Knowledge Graphs: a cascade of aliases, embeddings, and string fallback.

Strata

2008–2013
Shipping the social web PHP, game backends, 7×24 on-call Renren · Kabam (founding team, Beijing) · Zynga
2013–2018
Architecture at scale distributed rebuilds, throughput ×3 on the core ticketing system Meili Group · Baidu Lottery (dept. of 55)
2018–2023
Leading, then founding 150 engineers · K8s platform · H&M, Bosch, Ecco Squirrel PinPin (VP Eng) · Mojie ZhiXin (Founder/CTO)
2023–now
Canada. Back to the keyboard. edge-native commerce, AI agent tooling, on-device inference Conestoga GC '24 · NAU M.S. '25 · VantageCraft (consulting)

Now

Waterloo, Ontario. I run VantageCraft, an independent consulting practice. Recent work includes modernizing an insurance back office in Java/Spring and building a lab-services commerce platform on Cloudflare Workers and D1. I'm open to staff-plus engineering and technical-leadership roles in Canada.

Next on this blog: hybrid retrieval that fuses FTS5 BM25 with graph traversal, and an async ingestion queue built in plain SQLite.

Built with Astro, deployed at the edge, written by a person.