What's actually happening in the lab.

Local LLM Matrix v0.38.0 is out. Some screens had become overloaded, so there's now a calm Normal view for the actual decision and an Expert view with everything else. Plus one real new feature, the Matrix Tuner, which suggests starting settings and builds an Ollama Modelfile without touching the original model.

Picking Market & World Intelligence back up this week after the pause, and the first thing I found wasn't a new feature, it was a bug that had been sitting there quietly for a while. A field-name mismatch left the conflict risk score always empty, no crash, no log entry, just a health-coverage badge quietly showing the wrong status until I noticed.

Been quiet here for a few weeks, less time for AI Control Center, Local LLM Matrix or this Journal than usual. Nothing broke in the meantime, nothing got fixed either. Picking it back up now, starting with an actual update to Local LLM Matrix.

Second thread alongside the voice work: reaching AI Control Center, and Jarvis by extension, from outside the house. The Mac that runs everything sits at home, all I carry around is an iPhone 15 Pro. No approach settled yet, just the problem stated plainly.

Setting up Jarvis as a spoken layer on top of AI Control Center, so I can say what I want instead of clicking through panels. Getting a command to land on the correct mode instead of typing itself into whatever chat happens to be open is the actual work. Not solved yet.

Market & World Intelligence is basically feature-complete now, geopolitics, finance, climate, cyber, maritime and aviation all covered. Opened it this weekend and realized more data had made it harder to actually decide anything, not easier. No fix yet, just the diagnosis and a plan to show less by default without cutting capability.

Tested Obsidian's newer Bases feature against my test vault and found it silently dropping notes from filtered views. Three older notes had their tags property typed as text instead of a list, which Bases quietly excluded with no error. Fixed the template and converted the notes by hand.

Realized my Obsidian vault has never had a real backup, iCloud sync doesn't count. Wrote a nightly git backup script, then found .obsidian/workspace.json rewriting itself on every app open and burying every real diff. Fixed with one line in .gitignore.

Caught a coding agent twice in a couple of weeks slipping a fabricated reference into a solid piece of writing, a changelog line citing an issue number from a tracker that doesn't exist, then a test run claiming 38 seconds when the real CI log said 6 minutes. Now no number or reference goes out without me pointing at the real source first.

Rebuilt the image side of AI Control Center into three clear sections and fixed a progress bar that used to fake percentages ComfyUI never actually sent.

Went looking for a prompt to quote in a write-up and found out Local LLM Matrix never saved it anywhere, the data model just had no field for it. Fixed in v0.37.3, along with a live control center, a reset button, Ollama auto-start and a switch to pipx for installs.

AI Control Center is the one item on the "soon" list that's actually further along than the label suggests. Today it caught two Ollama processes silently fighting over the same port.

Renamed a project note straight in Finder today instead of through Obsidian. Seemed faster. Wasn't.

Spent the evening testing whether Q5_K_M quantization of Qwen2.5 14B would run better than the Q4_K_M I've been using for a few weeks, on my 18GB MacBook Pro M3 Pro.

Back in an old note from a few weeks ago I promised myself one rule for the vault. Every new note gets a date and at least one tag.

Spent way too long this week trying to get a local 7B model to output clean JSON for a small script, no dedicated JSON mode available on that one.

Spent Friday evening running the same small refactor through two different coding agents to see how far apart they actually are in practice.

Started actually building Obsidian AI Vault this week instead of just talking about it. Search works now, indexing is still too slow.

Set up git for this site today, first time this project's had real version control.

Two weeks since launch. Four projects marked "soon" on the build log are still marked "soon."

Caught something today while cleaning up the build log: it still listed "Local LLM Scout" as something coming soon.

Out of curiosity I pointed a coding agent at this site's repo and just said "clean up the CSS."

I tried to reorganize my notes vault this weekend.

Spent an evening this week convinced a 7B model was just bad at a summarization task.

Shared the link in a couple of places over the weekend and got the first outside eyes on the site.

goekmen.ai is live as of today.