great project, Fazle. there are definitely a lot of agent browser tools, but none of them are solving the core regression testing problems. This is def going deeper into this domain.

The weight_lbs to weight_kg example is a good one. Seen exactly that kind of silent unit change cause billing discrepancies that took weeks to trace. Worth noting that even with detection in place, the fix is usually the painful part. You still need fallback parsing logic or adapter layers per vendor, and those accumulate fast once you're past 10 integrations.

The 45 degree threshold is an interesting choice. Generous enough that normal mouse use passes but tight enough to catch most randomized sandbox scripts. Curious if anyone has tested whether Bezier curve interpolation in sandbox mouse emulation would consistently beat this.

this is some seriously thorough benchmarking work. the way you isolated each storage pattern and measured the overhead so precisely is impressive. benchmark #4 results are wild. mawk and nawk just completely falling apart on string concatenation while gawk barely breaks a sweat. the "structure penalty" finding is the kind of thing you only learn from actually measuring it. 5x to 8x more memory just for splitting fields vs storing raw lines. easy to overlook until it blows up in production. good stuff.

solid post marco. the 70% problem is very real. the junior dev question is something i think about a lot. if they never struggle through the hard parts, who's going to be the senior architect 5 years from now? and yeah on the PM point. as a technical founder i already do discovery, PRD, and then jump straight into building. that whole middle layer is getting squeezed hard.

this is cool.

Good one, tag #show-and-tell

cool idea. which models are you running locally? the quality gap between llama 3 8B and something like Claude is still pretty big for content gen. also curious about the hardware floor. not everyone has a GPU that can handle decent sized models at usable speed. any plans for a Linux build?

cool idea. running everything local with Ollama is the right call if you want zero ongoing costs. few questions though: which models are you running? because the quality gap between something like llama 3 8B and a hosted model is still pretty big for content generation. local is free but if the output needs heavy editing every time, the time cost adds up. what's the hardware floor? not everyone has a GPU that can run decent sized models at a usable speed. how are you handling context? content generation usually needs longer context windows and that's where smaller local models start to struggle. the "no API cost" angle is appealing but i'd be curious to see some actual output samples compared to something like Claude or GPT-4. free doesn't matter much if the content isn't usable without rewriting half of it. is this Windows only or are you planning cross-platform? would definitely try it out if there's a Linux build.

honest question - what does this actually do better than something like Strapi or Directus? CodeIgniter 4 is fine but the PHP CMS space is already crowded. "fast and lightweight" is what every CMS says. the real questions are: what's the plugin/module ecosystem look like? because building everything from scratch gets old fast on client projects. how does content modeling work? is it code-based or does it have a UI for defining content types? any headless/API mode or is it tightly coupled to the frontend? also "full open-source code available on March 17" means i can't actually look at the codebase yet. that makes it hard to evaluate. the demo site loads quick though, i'll give you that. not trying to be negative, genuinely curious where this fits. if you're targeting agencies building client sites, the comparison isn't WordPress. it's the modern headless CMS tools that already have API-first workflows and content previews figured out. what's your actual differentiator beyond being lightweight?