Category
Short engineering notes: libraries, frameworks, and lessons from building.
12 articles
Why I started moving away from Cursor and Claude Code subscriptions toward API-based AI coding workflows using DeepSeek, Kimi, and OpenCode — and what changed in how I think about cost and model choice.
Getting "The API Key appears to be invalid or may have expired" when using Kimi K2.5 or K2.6 in OpenCode? Here is the exact fix that worked for me using Moonshot AI instead of the Kimi Coding provider.
A breakdown of the relationship between pdfjs-dist and react-pdf, and why understanding the dependency tree matters.
Both extract text from PDFs; one is a full engine with layout and workers, the other is a quick dump. I picked sides for an AI reading assistant.
Indexes speed reads until they slow writes and eat RAM. I finally internalized ‘frequently queried’ as a design choice, not a checklist.
Parallelizing everything is tempting. Some pipelines only work in order — payments and balances taught me that the hard way.
Regenerating slides with pptxgenjs on every request was slow. Hashing the input and reusing the stored file when nothing changed made the API feel instant.
Next only picks up certain names for social images. I used og-image.jpg and wondered why previews broke until I read the convention list.
TTL is the right hammer for expiring documents, but storage, indexes, and delete load still count on small clusters like Atlas M0.
Client-side navigation can leave you halfway down the new page. A small effect tied to the pathname fixes it when you want a fresh scroll each time.
Each tool writes its own lockfile. Mix them and you get phantom bugs. Here is the cleanup ritual I use when I change my mind mid-project.
Copy-pasting the init command from the UI failed on Windows because & splits commands. Triple-quoting the URL fixed it.