Notes on engineering, architecture, and shipping.
Lumorem uses an LLM to judge performance findings. Pinning the first answer made audits reproducible, and hid the real question: how do I know the first answer is right? What the research says, and what I rebuilt.
Some waits go unnoticed, others make you close the app. What research says about how we experience waiting, with no technical background needed.
Lumorem ranks performance findings by user impact. The scoring formula started from the numbers that circulate about web performance. Tracing them to their sources reshaped it.
The audit pipeline for Lumorem is AST flags feeding an LLM. Half the work is filtering noise an off-the-shelf linter would have shipped. The other half is originating signal no linter produces.
A tour of the mechanisms keeping your photos private: EXIF stripping, RLS, signed URLs, etc...
Direct browser-to-storage uploads push the security boundary out of application code. Where the boundary sits instead on MirrAI.
Why MirrAI moved off SSE and the two architectures I built for streaming progress without holding a connection.