Things I've built, tested, and had to fix, mostly in security and self-hosting.
-
I ported Slay the Spire 2 to iOS (I typed, Claude ported)
read →My copy of Slay the Spire 2 now runs on an iPhone and runs better on my Android tablet. I would love to tell you I engineered this. What I actually did was own the game, own the devices, type requests like “can you make an apk release on my fork” into Claude Code, and paste Steam Guard codes into GitHub issues. Claude wrote the code. I’m told this is called “leadership”.
-
Dart security rules for SonarQube Community, with OpenGrep
read →Sonar has its own Dart and Flutter analyzer, but it ships in Developer Edition and up. On the free Community Build you get Dart through sonar-flutter, a community plugin. It adds the
dartlanguage and importsdart analyzeresults. Those are lints. A Flutter app that turns off TLS certificate checks scans clean. -
Building a SAST setup on top of SonarQube
read →I wanted a SAST setup for a set of mobile and web repos, mostly Flutter apps and a few web apps. The developers already work in SonarQube. They review hotspots there and their builds wait on its quality gate. So I wanted the security findings in SonarQube too, not in a separate dashboard nobody opens.
-
What happens when I revoke an agent's access
read →I want an agent to compare migration files with staging and report what’s missing. For that job it needs database access. The hard part is stopping it: once a task is revoked, can work already admitted still change the database? I have not tested this yet. This post is the design question and a small lab plan.
-
Can my local 8B do what Jev does
read →TypeSafe’s Jev showed up in my feed twice in one day. Both demos used it to pick from a fixed set of answers, not to chat. One sorted 1,018 research papers into 24 topics for eight cents, at a median 256 ms per paper. The other ran a flight search in seven seconds for $0.0039.
-
4B vs 23B on the 3060
read →My GPU box runs GLM-4.7-Flash as its local model. It’s a 23B mixture of experts with about 3B active per token, pruned and quantised to fit on a 12 GB RTX 3060. It writes well. The pruning and quantisation also make it unreliable on some facts.
no posts found for this tag.