About
Building infrastructure for reliable LLM systems.
Salvatan
I've been working with LLMs since the GPT-3 days, starting with content generation tools and eventually moving into production systems at early-stage AI companies.
The problem I kept hitting: prompts change constantly, but we treat them like config strings. No versioning, no systematic testing, no rollback strategy. When something breaks in production, you're scrambling through git history and Slack messages trying to figure out what changed.
PromptOps is the internal tooling I built at my last company, now available as a platform. It's designed for teams who need to ship LLM features reliably and iterate quickly without breaking things.
I'm based in Berlin, working on making prompt engineering less chaotic and more systematic.
Timeline
Started exploring GPT-3 for content generation tools
Joined early AI startup as prompt engineer
Built internal eval infrastructure for production LLM workflows
Moved to Berlin. Started PromptOps to make evals accessible
Open beta launch. Token launch planned for Q1
Principles
Reliability over hype
LLM systems are non-deterministic. Rigorous testing and versioning make them trustworthy.
Transparency by default
Open eval results, public roadmap, on-chain treasury. No black boxes.
Tooling for everyone
Infrastructure should not require a team of ML engineers. Make it accessible.
Why Berlin?
Berlin's AI scene is pragmatic and less hype-driven than SF. People are building actual products, not pitching visions. The focus is on making things work for real users, with strong attention to privacy and compliance (GDPR-native thinking).
It's also a great place to build infrastructure: lower costs, strong technical talent, and a timezone that covers both EU and partial US overlap.