Ruo Lee — Notes

I tried to steer users with a short survey toward re-selecting by rewriting a banner, and lost. Then I put the recommended action in the primary button's place, adopted a set of interaction rules for the app, and looked at the screens again through them.

The title of this post is something I said while we were polishing the banner text on the survey screen. People don't read those things, they just tap next out of habit. After saying it, instead of polishing the text further, I moved the…

Weekend signups dropped and I went in expecting an outage. It was the seasonal downswing of the whole market, and once I measured the cohorts it turned out this was never a retention business.

Weekend signups dropped noticeably. I went in assuming something had broken, either in the product or in search traffic, and the answer was neither. The market itself was on its way down for the season, and while I was in there measuring…

I decided when to show the inset map by running all 155 stops. Brussels was the single exception, and the 104px size stayed an eyeball call to the end.

My round-the-world site keeps a small globe in the bottom right corner. It tells you where in the world the city you are looking at sits. In Brazil it was useless. When you look at a city the camera comes in close. At that height the screen…

Turning on R8 as Play Console required stripped the Cordova payment plugin registry. In-app payments sat at zero for ten days; one keep rule fixed it.

Most of the revenue on the OPIc speaking-practice app I run comes from store in-app purchases. Of 54 payments across the app's whole life, 40 of them — 74% — are the app. There is a payment path on the web too, but the app is the one that…

Re-entering the start function across its await threw away the initialisation segment. I blamed iOS twice and was wrong twice; printing the first bytes of the upload settled it.

The speaking-practice app I run records a user's answer, sends it to a model and returns an assessment. If recording doesn't work, there is no product. On 8 September a new user failed analysis five times in a row, starting three minutes…

The server runs on UTC and the user's day starts in KST. Six bugs with nothing in common on the surface shared one root: two reference points mixed inside a single function.

Six bugs were fixed within one week. The symptoms had nothing in common. The weekly ranking didn't refresh on Monday morning. Review cards didn't appear on their due date. Credits refilled without limit. The admin dashboard and the admin…

I measured the scale before picking a countermeasure. At $0.27 the answer was to do nothing, and to remove one existing defence that was causing more harm than it prevented.

The app I run gives free credits on signup. Credits buy an AI analysis, and every analysis costs real money. When users surged, the expected question followed. Is someone creating multiple accounts just to harvest the free credits? There…

Neither a dependency nor a fork worked, so we vendored. The price is a discipline: no formatter over that tree, commits kept separate, and what we cannot fix written down.

The storage and retrieval engine behind the memory service wasn't written here. It came from an open-source library. But as described in the first post, that library's extraction prompt had to be changed — Korean conversations were being…

The deploy succeeded, the route was registered, and calls still returned 404. FastAPI's Mount matches on prefix and takes the first hit; allowing a slash in the alias was the cause.

A customer reported that calls to a newly created agent always returned 404. Everything we could check looked fine. The console showed the agent as deployed, the API access toggle was on, and the registration API had returned 200. Other…

Six defects that never raised an error. All of them used one empty array for both "nothing there" and "could not fetch"; four were caught before deploy.

The report came in like this. No error logs. No 5xx. The response was a 200 with an empty array where the memories should have been. Calling it repeatedly during the day never reproduced anything. Once I found the cause, I realised there…

I started at 0.9, then measured real sentence pairs through the production embedding model and landed on 0.85. The two directions of failure cost different amounts, so the slack went to one side; on the search path I dropped the threshold entirely.

The last post was about the memory service attached to an agent. This one is about a single number buried inside it. A user says "I'm moving in December" and the model writes that down. But last month it already stored "moving in November".…

The extraction prompt carried 14 English few-shot examples. No placement of an instruction changed anything; cutting to three examples with one Korean pair did.

I was building long-term memory for an agent. Facts a user drops during a conversation get extracted, stored, and recalled in later sessions. One day I opened the store and read this: The conversation had been in Korean from beginning to…

Naming the Redis channel after the request ID gave a nested call two subscribers. Five tokens arrived as 195, and the same cause came back weeks later on parallel calls.

Code written by a user runs inside an agent. That code can call another agent. When it does, the answer comes back token by token, and those tokens travel on to the browser. I ran a single line: print("Hello World"). The model produced five…

The same ten lines were rewritten five times in 24 hours. The version that lost nothing was the first one, deleted after 17 minutes — the code lived inside a generated template string.

There is one promise we made when integrating external model providers: whichever provider is used, the shape of the response a client receives has to be the same. Provider SSE payloads differ from each other. Field names differ, nesting…

Turning user Python into a live API without a restart means the code has to be a real file on disk. Three classes of bug came out of that, and three patches in I changed direction.

What this engine does fits in one sentence. It takes a Python file a user wrote and turns it into a live API without restarting the server. That left no choice about where state lives. You can't put code in a database. importlib doesn't…