A live update
on work and life.

I'm building a Japanese-learning app for myself, and the comprehension quiz taught me a lesson about how easy it is to accidentally build something that doesn't test what you think it tests. I had a multiple-choice check where you read an English cue and pick the right Japanese tiles. Turned out you could get 88% of the items right just from the English options alone, without reading any Japanese — the tile set was leaking the answer through wording. Flipping the direction (Japanese cue, English tiles) and writing a test that measures "solvable from English alone" as its own metric got that number to zero.

Furigana turned out to be more interesting than I expected. The naive approach — print the reading above the word — breaks the moment a word mixes kanji and kana, like 好き, where the き is already spelled out and doesn't need a reading repeated over it. I ended up building a small alignment step that strips the shared kana from both ends and only puts ruby over the part that's actually kanji, checked in both directions so a bad alignment fails loud instead of silently mislabeling a word.

I also spent a chunk of time chasing down a phantom bug on one of my home GPU boxes: a speech-to-text service returning 502s for no obvious reason. It wasn't the network — it was VRAM contention with an unrelated model sharing the same 6GB card. The health check reported the STT service as healthy because the "list models" endpoint doesn't touch the GPU, so it looked fine right up until it tried to actually transcribe something and ran out of memory. Fixed by working out the KV-cache math for the other model and trimming its context size until both fit with headroom to spare.

On the image-generation side, I learned that style-matching against 200+ existing images is mostly about restraint. Newer models kept "improving" on the reference style with gradients and reflections that didn't match — proof was a plain side-by-side against the originals, not a vibe check. And banning an object outright (like a prop) just pushed the model toward describing it a different way; the fix was closer to writing around it than negating it. Small, unglamorous findings, but they're the kind that only show up once you've generated a few hundred images and started comparing them side by side instead of trusting the first result.