Jieun
deserve what you want
14 posts
-
Eight Cheap Calls Beat One Good One — Why I Removed the Model That Wrote Better
I ran both engines on the same topics for ten days. Codex won on a close read of the output, and Codex is the one I deleted. No prompt could push either model past a length ceiling; what broke it was splitting the call, not swapping the model. Once the shape became many short calls, per-call latency decided everything. Grok's quality gap got closed with 4,000 characters of source and a code-level diff of every number.
-
Keyword Discovery: 25 Seconds Down to 0.2, by Calling Search Autocomplete
We mined Naver Q&A threads for keyword fragments and cleaned them up with two LLM calls. Replacing that with Google and Naver autocomplete took the button press from 25.1s to 0.2s and removed the LLM entirely. Google told people to stop using this endpoint in 2015, so there was something to settle before performance.
-
How Much Do You Open Up to an Outside Collaborator? Designing Isolated Dev Access Behind a BFF
The platform has exactly one door open to the internet, and every credential that touches data stays behind it. When a designer asked to run the UI locally, that collided with the design. I wrote down what to open and what to lock first, then issued an SSH key that can forward one port and nothing else.
-
Connecting a Custom Domain with One Click — a Pipeline Built Around 'You Can't Speed Up DNS Propagation'
I built a feature to buy a domain and attach it to a blog. Bought four with the same code; all four failed differently. One cause: we ignored a wait we can't shorten (DNS propagation) and tried to do everything synchronously. Once we accepted that wait as a given, the pipeline settled into one synchronous stage plus two worker stages.
-
Taking Caddy Out of the Deploy Path — the Structural Problem Was Touching a SPOF Automatically
Every deploy took the whole stack down for 60–100 seconds. Blue-green removed most of it, but five seconds stayed at the reverse proxy — and the real cost of those five seconds was not five seconds. It was the variance: five seconds, or indefinitely. So we took it out of the deploy path entirely.
-
[Threads Marketing Agent, Part 3] Multi-Account Engagement Automation with an Anti-Detect Browser — Brain/Hands Split, CDP Driving, Detection-Aware Pacing
Part 3 of an engineering log on building a Threads marketing agent: automating engagement (likes/comments). Isolating sub-accounts with an AdsPower anti-detect browser, connecting a VPS (brain) and a Mac mini (hands) over a queue, driving the browser via CDP, and pacing it with a detection-aware random schedule.
-
[Threads Marketing Agent, Part 2] Designing a Generation Pipeline — A 4-Axis Content Model, Queue-Based Scheduling, 3-Layer Attribution
Part 2 of an engineering log on building a Threads marketing agent: a content model decomposed into four axes, an idempotent publish worker, and 3-layer attribution where 90% of clicks were bots.
-
[Threads Marketing Agent, Part 1] Designing a Crawler That Beats Three Tiers of Anti-Bot — From TLS Fingerprints to CDP Detection
Part 1 of an engineering log on building a Threads marketing agent: the bottleneck was never LLM generation — it was crawling. Tiering the strategy by which signal the anti-bot actually inspects.
-
Scaling QPS 10x in Production: 3 Strategies That Actually Move the Needle
"Make it faster" sounds vague. It isn't — it comes down to the same three moves: caching, database optimization, and pulling non-essential work out of the request.
-
[Messenger Load-Testing, Part 3] FCM to a Queue, Swallowed Failures to Sentry — Cleaning Up the Message-Send Handler
Part 3 of load-testing a messenger: pulling FCM push out of the message-send handler into a dedicated worker queue, and surfacing the failures that used to vanish into empty catches through Sentry. Two axes — isolating external I/O and handling error logs.
-
[Messenger Load-Testing, Part 2] There Was a Deadlock Hiding in AI-Generated Code — How We Replaced `unreadCount` with `lastReadAt`
42.8% HTTP error rate, 9 minutes into a load test, at 12% of the target TPS. The culprit was a single updateMany line written during AI-assisted coding.
-
[Messenger Load-Testing, Part 1] Building a Reusable Load-Test Environment — Make-based IaC, Observability, and Auth Seeding
How many TPS can a messenger's send API survive, and where exactly does it break? A 7-node mirror of production, rebuilt on demand from a single Makefile target.
-
Why You Should Never Use KeyDB as a Pub/Sub Broker for Centrifugo
KeyDB's active-replica mode looked perfect for a symmetric dual-stack chat at 100K WebSocket connections. Then I found five years of hang and deadlock issues, most still open.
-
Building a Crash-Resilient, End-to-End Meeting Recording Pipeline in the Browser
A one-hour browser recording could push the JS heap past 300 MB, and a tab crash meant losing everything. Rebuilding it as a six-stage pipeline with five-minute savepoints.
