Built by Ghiles, alone, on purpose.

I’m Ghiles, 24, an AI engineer in Stockholm. I work as a consultant, mostly on AI features for SaaS products. Over the last two years I shipped LLM-powered things on more than half a dozen client projects: chat, document summarization, classification, agents.

Every single one ran into the same problem. Nobody could tell, in real time, which customers were quietly unprofitable.

Why this exists

The pattern is always the same. The OpenAI bill arrives at the end of the month. You compare it to revenue and the gross margin looks fine in aggregate. Then you dig deeper and find 5 percent of users are burning the other 95 percent’s profit. By then it’s already happened.

I tried every existing tool: Helicone, Langfuse, OpenLLMetry, custom dashboards built in spare-time weekends. Every one of them gives you traces or token counts. None of them give you margin per user. I needed numbers I could actually act on, not after-the-fact telemetry.

So I built Weckr. It is the tool I wished existed on every client project. Wrap your LLM client in two lines, see per-user cost vs revenue in real time, set spending caps that block or downgrade before the bill arrives. The whole thing exists because I got tired of doing the math after the fact.

Building in public

I document the build at @cyyylas: shipping logs, dumb mistakes, what works for an AI SaaS that isn’t VC-funded, the honest economics behind a tool that itself wraps LLM calls.

Want to talk?

If you’re building something AI-shaped and want to compare notes, or you’ve hit the same margin problem and want to argue about pricing models, email hello@useweckr.com, DM me on any of the above, or open a discussion on the public SDKs repo.