Before rebuilding my main app I wanted numbers instead of folklore. Not synthetic hello-worlds the same five workloads my app actually does: render a DB-backed page, serve JSON, take events over We...

This is part four of four. Part three locked a recipe: detect three times, sample the text color from the brightest glyphs, neutralize strokes with a median patch, paint one constant rectangle at a...

This is part three of four. Part two ended with a new goal and three acceptance criteria: deterministic, auditable, honest. The watermark becomes one uniform translucent rectangle carrying its own ...

This is part two of four in the deployment tools series (part one: OpenShip). Where OpenShip wraps your box in a dashboard, a service catalog and a guided mail wizard, Dokku is the opposite bet: ev...

This is part two of four. Part one ended with a working pipeline that detected watermarks and let LaMa rebuild the background under them, and with three cracks in the approach: invented content, no...

My scraper pulls real estate listing images into MinIO. Every single one carries the host site’s watermark. The count was 113,998 when I first attacked this problem, and it has since grown to rough...

This is part one of a four part series on deployment management tools. I am putting four platforms through the same gauntlet on identical VirtualBox VMs Ubuntu Jammy, 2 vCPU, host-only networking, ...

This story got a deeper rewrite as a four part series with diagrams and code. It starts at How do we remove watermarks. Part one ended with a locked recipe: detect the watermark three times, s...

This story got a deeper rewrite as a four part series with diagrams and code. It starts at How do we remove watermarks. This is part one of two about cleaning watermarks off scraped real estat...

I used to think completion was about content. Then I pulled retention for a cohort. Most learners who stopped did not fail. They just went quiet after a due date passed. No nudge, no follow up. I h...