Dottxt —Making AI speak computers

When we first encountered .txt (operating as dottxt), we'll admit we were skeptical. Another AI startup promising to "fix" large language models? But after digging into what founders Rémi Louf, Dan Gerlanc, and Brandon Willard have built, we're genuinely impressed by their approach to a real problem that's been driving developers crazy.

The team's origin story resonates with anyone who's wrestled with LLMs in production. They were at Normal Computing, trying to extract structured data from GPT-4, and kept hitting the same wall—unreliable outputs that required constant babysitting. Instead of accepting this as "just how AI works," they decided to engineer their way out of the problem.

The Technical Breakthrough That Actually Works

What caught our attention isn't just the promise of "100% valid JSON outputs"—it's that they've actually delivered on it. Their open-source Outlines library has been downloaded over 3 million times, and here's the kicker: OpenAI and Cohere are using their technology. When the giants in the space adopt your solution, you're probably onto something significant.

The core insight is elegant in its simplicity. Instead of hoping LLMs will produce structured data correctly, dottxt ensures they can't produce anything else. It's like putting guardrails on a highway—the model can still drive wherever it wants, but it can't veer off into invalid territory.

Why Enterprise Teams Are Paying Attention

We've seen plenty of AI tools that work great in demos but fall apart in real-world applications. What's different here is that dottxt is solving the reliability gap that keeps enterprises from fully embracing LLMs. When you're processing thousands of CVs or generating compliance reports, "mostly accurate" isn't good enough—you need guarantees.

The fact that they've raised $11.9 million from serious VCs like EQT Ventures and Elaia in just over a year tells us the market validation is real. These aren't vanity metrics—enterprise customers are willing to pay for structured, dependable AI outputs.

The Open-Source Strategy That's Actually Smart

Here's what we find refreshing: dottxt isn't doing the typical "open-source for marketing" play. They've built genuine value in their Outlines library while reserving their optimized platform and enterprise features for paying customers. It's a thoughtful balance that builds developer trust without giving away the farm.

For teams already struggling with unreliable LLM outputs in production, dottxt represents a practical path forward. They're not trying to reinvent AI—they're making existing AI actually usable for serious applications. Sometimes the best innovation is just making things work reliably.