Real-time video · product discovery · education
Spanish for Connection
I built Spanish for Connection around a problem I understood personally: it is possible to know a lot of Spanish and still be unable to use it when a real human starts talking back.
I knew Spanish. I couldn't use Spanish.
After six years of school Spanish and 4+ college semesters, I could write an essay about El Greco or Gabriel García Márquez. Then I got to Medellín and sometimes couldn't order a tintico on the street without people stepping in to offer translation.
That was maddening because I knew the language. What I did not have was reliable access to it under real conversational pressure.
I could write about García Márquez. I couldn't reliably order coffee.
The constraint was not more vocabulary.
School had trained me to care about accuracy. Accuracy matters. But live conversation also requires reps, recovery, and tolerance for imperfect real-time output. Miss one word, start translating in your head, fall behind, get embarrassed, go silent, and suddenly the other person switches to English.
That led to a different product question: instead of teaching more Spanish first, could I train the skill of staying in the conversation?
What we are teaching is how to stay in the conversation: slow the exchange down, remain in Spanish, show what landed, locate what did not, and keep going without treating imperfect comprehension as failure.
Conversation repair is an interface.
A generic “Repite, por favor” is not wrong Spanish. It is a low-information repair move. The other speaker still has to guess whether the problem was speed, vocabulary, one phrase, the whole sentence, or even whether you understood the intent.
The stronger move is to expose partial state: identify whether you heard a question, statement, request, or suggestion; say what you did understand; point to where you lost the thread; then ask the other person to repair that specific gap.
- Buy time without disappearing. Stay in Spanish instead of going silent.
- Reduce the search space. “It sounded like you were asking me…” is much easier to repair than “I didn't understand.”
- Give the other person diagnostic information. They can adjust speed, vocabulary, or phrasing to the part that actually failed.
- Make imperfection collaborative. The learner is still participating rather than waiting to be rescued.
The live product turns speaking into a game.
The live experience grew into an Around-the-Horn-style speaking game: I host, students are contestants, and the interaction is designed to create lots of short speaking turns without making every turn feel like a surprise oral exam.
- Speaking earns points and keeps the game moving.
- Silence costs a point and advances the turn rather than letting a freeze stretch indefinitely.
- A visible queue shows who is coming up next, reducing the anxiety of being called on unexpectedly.
- I can award subjective points for effort and courage, not only linguistic accuracy.
The design goal is not to remove pressure. It is to create enough structure that students can practice functioning inside it.
The real-time video platform is built on LiveKit + Next.js. The product pairs a weekly live speaking session with a broader practice system designed around high-volume output and fast feedback.
The second layer is daily speaking practice.
The program concept pairs the weekly live session with short asynchronous prompt-and-response reps. The goal is to make speaking a frequent behavior rather than something that happens for one scheduled hour each week.
I also prototyped pieces of a more unusual pronunciation loop using voice synthesis: model the learner's own voice, generate a personalized target, and let the learner practice toward that target rather than toward a generic reference speaker. The idea is to make “better pronunciation” feel like a reachable version of your voice, not impersonation.
The broader asynchronous practice and scoring layer is still in development. Pieces are built; the complete daily practice loop is not.
The market test
I did not leave the idea as a private prototype. I sourced a school directly, met with a board director and Spanish staff, and in the first stakeholder meeting received verbal agreement to a free pilot.
The pilot has not yet run. I therefore do not claim student outcomes from it.
What changed because I built it
- A personal frustration became a behavioral hypothesis: language knowledge and speaking performance are related, but they are not the same constraint.
- The interaction model treats anxiety as a product variable: visible turns, fast recovery, frequent reps, and rewards for effort as well as correctness.
- Building the real-time experience surfaced design questions that a curriculum document would not.
- Direct stakeholder discovery moved the work from a private prototype to a proposed school pilot.
- The unfinished async layer stays labeled unfinished rather than being promoted by aspiration.