Cuéntame Ella

The idea started with a real observation: family caregivers in Latino immigrant communities spend hours as informal medical interpreters in situations they were never prepared for. Waiting rooms, insurance calls, hospital discharge conversations. The family member who speaks the most English ends up in the middle, trying to translate medical terminology they do not understand, in real time, without making a mistake.

They brought this to the first session already knowing what they wanted to build. A voice and text assistant in Spanish that could translate in real time during a medical appointment, take notes, translate prescriptions, generate medication reminders, and store a summary of the conversation for 42 to 72 hours so family members could review it later. The design instinct was right. The technical path to get there was harder.

The first challenge was the platform. The original plan was WhatsApp because that is where people already are, especially older adults who would never download a new app. WhatsApp turned out to be the wrong starting point because of how difficult and expensive it is to get business API access. The team switched to Twilio, which allowed the same SMS and call-based interface without the verification barrier. Platform agnostic. Works on any phone.

The second challenge was language. Spanish is not one language in practice. During testing, the team ran into the word "buche." In some dialects it means the throat. In others it refers to the stomach. In a medical context, a patient describing pain in their "buche" could be pointing at two completely different parts of the body. If the translation tool defaults to the standard dictionary definition and gets it wrong, the doctor is treating the wrong symptom. The team built an edge cases document to map out exactly this kind of problem, word by word, dialect by dialect. The list got long quickly.

The third challenge was HIPAA. The tool stores summaries of medical conversations. That storage, even for 42 to 72 hours, puts the project in proximity to federal health privacy law. The decision in the early sessions was to build the prototype first and treat compliance as a later layer, but the team was clear-eyed about the fact that this is a problem that has to be solved before anything goes live with real patients.

The deeper technical testing revealed more problems. When the tool was tested in doctor-patient role play, the agent repeated everything in both languages to confirm it heard correctly, which became disruptive. It translated literally when it needed to infer. When the patient said something coloquial, the tool produced a word-for-word English rendering that missed the meaning entirely. Those are not small bugs. They are the core of what the tool is supposed to do.