How Speak uses Raindrop to build better agents for 15 million users

Speak is an AI-powered language learning app that helps people become more fluent in any second language. Today, Speak serves well over 15 million users across more than 40 countries, with their AI tutor at the center of the experience.

Surfacing the Long Tail

As Speak scaled their conversational tutor to millions of learners, the hardest problems were the ones the team could not see. Failures happened quietly in production, and users would get frustrated and churn before anyone on the team knew something had gone wrong. Once Speak implemented Raindrop, the team finally saw how people were actually using the tutor, surfacing a long tail of edge-case interactions and failures that never appeared in traditional logs or dashboards but quietly eroded the experience.

"Once we implemented Raindrop, it was frankly a little bit shocking the extent to which people were using and talking with the tutor in these super unexpected ways. There was this huge long tail of issues."

Andrew Hsu

Co-Founder & CTO, Speak

Protecting a High-Trust Relationship

Language learning runs on trust. Learners expect that everything their tutor teaches them is correct, so it is critical to Speak that everything the system says is accurate and trustworthy. When that trust breaks, the damage to the relationship is real.

To stay ahead of that, Speak runs a daily Slack channel where Raindrop posts what amounts to a daily incident report of everything going wrong.

"We have this daily Slack channel where Raindrop sends us essentially a daily incident report of all the stuff that is going wrong. It’s a little stressful because it is such a visceral representation of all these struggles users are having, and it’s also super motivational to fix it all."

Andrew Hsu

Co-Founder & CTO, Speak

A Standard Part of Every Launch

Raindrop is now a default part of how Speak ships. Whenever the team builds a new conversational agent, they start by defining what good and bad signals look like and what the product is really trying to do, then wire Raindrop up before launch and monitor it closely.

"We always try to think about what are the good signals, what are the bad signals, what is the product intent, what do we care about, and we wire up Raindrop. It’s a standard part of our process now, because the signal we get from it is not really possible to get any other way."

Andrew Hsu

Co-Founder & CTO, Speak

Toward a Self-Healing Loop

Because Raindrop captures production signal and exposes it through its MCP server, a coding agent can pull the full context of a failing interaction, reason about the root cause, draft a fix to the prompt, and open a pull request on its own. Raindrop closes the loop from production issue to proposed fix, so the team spends its time reviewing changes instead of hunting for problems.

That foundation is what frees Speak to keep building toward its larger goal: a language tutoring super agent. One tutor you can talk with and text with, that has a full understanding of you as a learner and brings you to your fluency goal as efficiently as possible.