y30
Co-Founder & CPTO · since 2026
Voice AI for older adults and their caregivers. The bar I held it to: not "did it answer." "Would grandma keep talking?" I define problems with caregivers and clinicians, write the specs, design the IA and interactions, and ship end to end: Python, LiveKit, React, Supabase. Years in the unglamorous parts of healthtech: call centers, care-coordination handoffs. The places where the gap between the system and the actual person is hardest to ignore. y30 closes it: voice that earns trust, and every conversation becomes structured signal a caregiver can act on.
Healthcare is the thread. Voice is the medium. Dignity is the constraint.
The problem
Caregivers and clinicians are flying blind between appointments. A senior can decline significantly in days, but the system only finds out at the next visit, if at all. The signal exists; it just has no infrastructure to surface it.
What I built
- Made it slower, on purpose Most pipelines fire at 200–400ms. For an 80-year-old, that's not fast. It's rude. y30 uses a 3.2-second patience window with a contextual turn-detector that holds longer when a sentence isn't finished, and I cover the wait with immediate non-verbal feedback so it reads as listening, not lag. One call wearing three hats: the product judgment to not rush them, the interaction design to make waiting feel like care, and the engineering to hold the turn.
- An interface an 80-year-old never has to learn One tap to start. One task per screen. Tap targets sized for unsteady hands, type for tired eyes, and a live audio indicator so the person always knows it's listening. No AI sparkle, no chat-bubble clutter, nothing to figure out. The hard part was everything I left out, so the user never feels tested.
- Deterministic engine, warm LLM delivery Flow control runs on a finite-state machine. The LLM only owns phrasing, warmth, and repair. About 50% of routing decisions resolve in 0ms, no model call. LLM consulted only on ambiguity (p95 1.5s). Flow is engineering; warmth is the model. They don't bleed.
- Safety that survives the model going down A rule-based detection floor fires independent of any LLM or feature flag. The classifier adds nuance on top, never as the sole line. A crisis response can't be contingent on an API being healthy. The voice session is sacred.
What I didn't build
Killed preemptive generation: it tested well but produced ghost speech fragments in real conversations. No diagnosis, treatment plans, or financial advice, by design. Phase 1 had one job: prove voice quality. Restraint was a feature.
1,389 commits · 127 test files · STT 4-deep · TTS 5-deep provider fallback
Every conversation auto-graded by an eval pipeline. Regressions file themselves as issues, metadata only, no PHI.
A real reminiscence call, and what it becomes.
An unscripted voice session: the agent waits through a long pause without cutting in, yields the instant it's interrupted, and softens its tone when the conversation turns to grief. Then the conversation becomes structured signal a caregiver can act on.
Voice AI
Elder Care
Human-AI Interaction
Interaction Design
Real-Time Systems
Accessibility
Product Engineering
