Engineer's field notes
I build frontend systems that hold up in production.
I'm Koh Hom (许峰), a software engineer with experience across Shopee, Ant Group, ByteDance, and TikTok. These are my field notes: case studies from real projects, and long-form writing on web performance, testing, and AI-assisted engineering.
- Writing since
- 2022
- Notes
- 31
- Case studies
- 5
- Languages
- EN / 中文
Selected work
All work-
Understanding ecommerce growth data from an engineering perspective
I recently moved to a new team working on ecommerce growth at TikTok Shop. I had not worked as a growth engineer before, so I used the team's existing projects and GPT to learn the area. The biggest thing I learned is…
- Decomposed GMV into an engineering funnel (UV x PDP arrival x add-to-cart x checkout x payment success x orders per buyer x AOV) so a drop can be traced to the step that caused it.
- Designed the app landing experiment around one primary metric, 7-day Web-assisted App paid GMV per eligible MWeb UV, so the result is attributable to a Web touchpoint, really transacts in the app, and is measured per unit of traffic.
- A/B testing
- deeplink
- btm_chain
- funnel analysis
- cohort retention analysis
9 min read
-
Designing an Operations Heartbeat System
A system design write-up on turning UI activity events into reliable agent state across compute windows, HCM state commits, downstream propagation, multi-region cutover, and service split boundaries.
- The HCM and Heartbeat path serves 47,149 agents, 7,000 skill groups, about 4,000 QPS, and 141 upstream services.
- Splitting TT from the main business line later let the old HCM path shed about 3,000 agents and about 2,000 QPS while TT heartbeat, status writes, and routing stayed closed inside TT.
- Redis
- MQ
- DBus
- binlog
- Elasticsearch
16 min read
-
Workspace v2 tab system performance: first load, hot switch, and background pressure
A performance write-up about separating first load, hot tab switch, background pressure, and the gates that kept false wins out.
- Cold tab switch p95 duration fell from 1829.8ms to 812.3ms (-55.6%), and p95 post-visible blocking fell from 1193.7ms to 8.7ms (-99.3%) after moving prewarm before the click.
- Scheduling route FMP improved from 14773ms to 11926ms (-2847ms / -19.3%) by discovering route chunks early instead of at 6785ms-15464ms.
- CDP Chrome
- React
- iframe sandbox
- requestIdleCallback
- WebSocket
8 min read
-
Workspace v2 tab system: browser tabs inside a workspace
How we made workspace tabs keep state, survive refresh, share real URLs, and stop hidden runtimes from leaking side effects.
- Tabs survive refresh and stay identical across windows: BFF stores the opened tab list, React Query makes the local window instant, BroadcastChannel makes other windows refetch.
- The address bar keeps real business URLs like /workspace/workstream/:id/..., so refresh, bookmarks, and shared links recover the business page instead of an orphan /tabs/:id route.
- React
- TypeScript
- React Query
- BroadcastChannel
- History API
9 min read
-
Automated AI performance optimization with a harness and goal-driven loop
What changed when I made an AI agent optimize performance against a harness instead of against its own guesses.
- Workstream FMP improved from 5089ms to 2519ms (-2570ms / -50.5%) after removing the startup payload, warming the list API, and moving notification/DnD/editor/actions after the final FMP.
- Report Center FMP improved from 10021ms to 6762ms (-3259ms / -32.5%) by moving notification work after FMP and fixing the prefetch body/cache-key alignment.
- TypeScript
- Service Worker
- GraphQL
- REST
- requestIdleCallback
6 min read