Why is there a minimum age to become president, but no maximum? At 34, you're supposedly too young to run a country, but at 80, it's perfectly fine? Where's the logic?

GraphCompose 2.4.1: ~17× faster rendering in a local PDF test with QR + Code 128. 29 → 1.7 ms. Memory allocated per document: 7 → 0.7 MB. Barcodes now render as vectors in PDF/PPTX. No intermediate PNG. GH: github.com/DemchaAV/GraphCom… New gallery: demchaav.github.io/GraphComp…

GraphCompose v2.4.0 is out. 159 commits since v2.3.0. New CV, invoice, proposal, receipt and rota templates, real letter spacing, better timelines/lists, pagination fixes, ATS-tested CV presets and a lot of work inside the layout engine. It is becoming much more than a Java PDF library. github.com/DemchaAV/GraphCom…

Time for some funn with GC 😅

Today I tried a Fable. 😅

What happened, Claude and Grok are gone?

Google launched Gemini Flash 3.8 today, with benchmarks showing promising results. It prioritizes cost and latency over being top-tier. compare all these things, it could be worth. And what I noticed from tests is that it's trying harder than the previous model. But I wouldn't say it made a difference. Will see, I have hope that they will make it and release a top-tier model, which is Gemini 4. Or is there no hope anymore? What do you think?

The interesting part of building a better agent harness is that improvements don’t only make strong models better. They start making weaker, cheaper models usable. Gemini used to struggle badly with my document workflow. Now even the released version can get to a workable result. If I can push that far enough, the next experiment is a local vision model like Qwen. The goal isn’t “which model is smartest?” It’s how much intelligence can move from the model into the workflow.

I have tried Fable 5.1 5 ours limits are gone in a second, and it says 25% les costs safe 🫠 #Fable5

New GraphCompose v2.3.0. And now it comes with knowledge for AI agents. Stop pretending you write everything manually. Especially when you use a young library that AI models barely know. The model doesn't need to guess anymore. github.com/DemchaAV/GraphCom…

GraphCompose is slowly moving beyond “someone downloaded it once.” Scarf currently classifies 97 companies: 41 investigating, 11 experimenting, and 3 in ongoing usage. Still early, but those 3 are probably the most interesting number here.

Why Harness Is Becoming One of the Key Parts of AI Development Models are already smart enough for a huge amount of real work. What is missing is the software layer around them — the one that defines how a task is executed, verified, corrected and finished. A look at why harness is becoming the interesting part of agentic AI. @grok demchaav.github.io/blog/post… #AI

graphcompose-ai-flow harness 😮 #Harness

GraphCompose AI Flow actually works. I gave Claude Code a CV screenshot and one sentence. 77 minutes later, it returned a Java GraphCompose template that was already close to production-ready. Yes, it’s slow. But the result is maintainable and testable native Java, not a mockup. I only described two visual issues in plain language, and the agent fixed them itself. That’s exactly the workflow I wanted to build. 💥💥 github.com/DemchaAV/graphcom…

Another budget app? Not another expense tracker. Most apps tell you where your money went. BudgetPocket helps you decide where this month’s money should go—before you spend it. No account. No cloud. Your data stays on your device. Preparing the Android release.

Yes, I know Google and their Gemini are behind. But 3.7 flash feels fresh and quite quick with a decent reply, even with image input. Maybe they will recover with their frontier model, to join that competition 😁

GraphCompose 2.2.0 is out. Java documents that read right to left: Hebrew and Arabic lay out, shape, join and mirror — in paragraphs and table cells — through PDF, PowerPoint and Word. Five script families bundled. One DSL, three backends. github.com/DemchaAV/GraphCom…

😅

graph-compose-fonts 1.1.0 is now published. It adds bundled font families for Arabic, Hebrew, Georgian, Armenian and Korean. This is part of the groundwork for GraphCompose 2.2.0, where text direction, BiDi layout and Arabic shaping are becoming part of the engine rather than PDF-specific workarounds. More soon.