Pavel Racu

Pavel builds things.

I run product at AWSsome, an AWS Marketplace integration platform. I also build solo projects and write about both — what shipped, what broke, what I learned.

Projects
AWSsome
AWS Marketplace integration platform. Co-sell, deal registration, order management, product listings — the operational layer ISVs need to sell on AWS.
Active
dotpm
An AI writing tool where the document is the interface. Type an instruction on any line, get the output in place. No sidebars, no chat panels.
Active
Flair
AI analytics for B2B sales coaching. Analyzed video calls to help reps improve and give management market insights. Raised €150k, grew to 7 people, sold to a US competitor after 2.5 years.
Exited
LocalZero
Browser-based PII detection and redaction. Drop a file, scan for sensitive data, redact it, export clean. Nothing uploads — everything runs client-side.
Paused
Writing
Three onboarding surprises nobody warned us about
We shipped onboarding for ISVs selling on AWS Marketplace. The hardest parts weren't technical — they were the assumptions we made about what "ready to sell" means.
254 documents in 15 weeks
Briefs, research reports, sprint reviews, hiring posts, strategy docs. All written through conversations with Claude. The writing part works. The glue between tools doesn't.
Concurrent agreements broke our assumptions
AWS Marketplace lets buyers stack multiple agreements on one product. Our order management assumed one buyer = one agreement. That assumption made it to production.
I researched 14 AI writing tools. None solved the input problem.
Every AI writing tool faces the same design question: where does the user give instructions, and where does the output appear? They all fall into 5 paradigms.
The vibe coding trap for internal tools
We used AI to build an internal tool in two days. It worked. Then someone needed to change the date format, and the whole team stared at code nobody understood.
We built 6 prototypes in one session
Not mockups. Working prototypes in a BlockNote editor. Each one taught us something concrete about where AI instructions should live inside a document.
Get in touch