By Ezequiel Saraceno,
AI Engineer Manager at Learnlight
Before I retrained as an AI engineer, I spent 15 years as a neuroscientist. First in my home country at the University of Buenos Aires, then at CNRS in France, where I completed two postdocs.
Learning a new discipline while building a life in a language that wasn’t my own shaped how I think about language learning more than anything else could have. So, when my team at Learnlight was given the task of building Fluency Lab, an AI-powered language practice partner, I came to it having lived through that experience myself.
The Language Learning Problem We Set Out To Solve
The question I kept coming back to was this: how do you make AI genuinely useful for learning?
Every major AI provider—OpenAI, Anthropic, Google—offers conversational AI that learners can engage with in seconds. You can get language practice out of any of them, if you know how to prompt it correctly. For Learnlight, the bar was higher than that.
Raising the bar meant taking our expert-built pedagogical content, with the clear learning goals, vocabulary frameworks, and carefully leveled grammar, then shaping it into an AI experience that felt natural to use. In practice, that required matching our existing methodology, which keeps learning sessions to just 10 to 15 minutes, and creating something accessible for learners who might never have spoken with an AI before. Above all, it meant keeping the AI grounded: no hallucinations and no sessions that felt busy yet delivered nothing.
How Fluency Lab Works
The core of Fluency Lab is a conversational practice loop. A learner chooses a topic, reviews vocabulary cards, and enters a conversation with the AI trainer. The AI’s job is to guide them naturally toward using that vocabulary in context, through a real conversation that creates genuine opportunities to apply the language. At the end of the session, the learner gets immediate feedback: what went well, what went wrong, and what a stronger version of a given sentence would look like.
The Challenges Of Building An AI Trainer
Large language models (LLM) are generative by design. They want to keep conversations moving, fill silences, explore tangents, and be helpful in every direction at once. For a general-purpose assistant, that quality is useful. For a 12-minute language learning session with a clear pedagogical goal, it creates a real challenge.
The work of building Fluency Lab was largely the work of building control. We needed something that felt like a real conversation: warm, natural, responsive, and disciplined enough to serve the learning objective. That meant significant prompt engineering and rigorous stress-testing before anything reached the learner.
My neuroscience background was more useful here than I expected. Understanding how memory consolidation works, how retrieval practice strengthens retention, and how cognitive load affects acquisition: all of that informed specific design decisions about session structure, feedback timing, and how the AI handles errors. Building tools intended to improve human cognition requires a working understanding of how cognition actually works. That should go without saying, but in this field, it often doesn’t.
Making AI Something People Come Back To
The technical architecture was a solvable problem. The one that occupied more of my thinking was adoption.
We’re at a particular moment in the AI adoption curve where usage is growing fast, but the technology is still moving faster than most people’s comfort level with it. I was very aware that if a learner tried Fluency Lab once, felt uncertain or confused, and closed the tab, we’d potentially lose them for good.
So the learner experience had to do two things simultaneously: be pedagogically effective, and feel safe. Those goals don’t always align. A session that challenges a learner appropriately might also feel uncomfortable. An AI that corrects mistakes needs to do so in a way that keeps someone engaged rather than shutting them down.
We spent a lot of time on how the AI communicates, how it handles unexpected responses, and how it keeps sessions feeling productive even when a learner is struggling. For L&D teams thinking about rollout, that design work is what determines whether a tool continues to be used after the first try. Adoption is very much a product design problem, and it needs to be treated as one from the start.
Why A Human-First Approach Matters
At Learnlight, we’re human-first, technology-enabled. I say that as the person building the technology.
Some companies in this space take an AI-first approach, and I understand the appeal. My view, shaped by my technical background and by how quickly things are evolving, is that we need to be careful. LLMs are extraordinary tools but they’re also still generating language based on probability, not thinking for themselves. That’s why the human in the loop matters.
Bringing Fluency Lab’s AI Trainer To Market
We presented Fluency Lab at a conference in France in February. There was a genuine buzz in the room, and clients were already asking how soon they could get the tool in front of their teams.
That reaction told us everything we needed to know : we’d solved the right problem.
—-
About the author
Ezequiel Saraceno is AI Engineering Manager at Learnlight, where he leads the team building Fluency Lab’s AI-powered language learning tools. Before moving into AI engineering, he spent 15 years as a neuroscientist.
