SeniOra

Challenges We Faced

The Reliability Problem

ElevenLabs rate limits hit us hard during testing. A failed TTS call meant an elder sitting by the phone hearing nothing. We built a graceful fallback system — every single ElevenLabs call is wrapped in try/catch, falling back to Twilio's native TTS. The call always completes.

Real-Time AI Evaluation During Live Calls

Assessment calls need Gemini to evaluate an elder's spoken answer while they're still on the phone — transcribe audio, judge correctness, and generate a spoken response, all within the timeout of a Twilio webhook. We parallelized audio pre-generation and optimized prompts to keep latency under the wire.

Making AI Sound Human

Early versions sounded robotic and clinical. Elderly test users hung up. We iterated heavily on Gemini prompts to generate scripts that are warm, patient, and conversational — like a caring grandchild, not a hospital system. ElevenLabs voice cloning let us go further: elders hear a familiar voice.

Designing for Users Who Can't Use Apps

Our end users don't have smartphones. Every interaction happens over a regular phone call — DTMF keypresses for input, voice for output. No app downloads. No account creation. Just answer the phone.

Multi-Language Support

Alzheimer's patients often revert to their first language. We built multi-language support (English, French) so the system speaks to elders in the language they're most comfortable with — critical for accurate cognitive assessment.

Built with Google Gemini, ElevenLabs, Twilio, and a lot of empathy.