Inspiration
What it does# Haven: A Voice That Doesn't Make You Feel Old
Inspiration
Haven started with my grandfather-in-law. He's 82, sharp, independent — and technology had quietly become a wall between him and the world instead of a bridge. Every new device promised to help him but actually asked more of him: remember a password, tap through five menus, read tiny text. The tools built to serve elderly people were, in practice, built for people who already understand computers.
I watched that erode his confidence in small ways I couldn't unsee. That's the moment Haven stopped being an idea and became something I had to build.
The mission behind it comes from Matthew 10:8 — "freely given, freely give." I believe AI should be an equalizer, not another gate that only opens for people who already have access, money, or technical fluency. Elderly users are one of the clearest examples of a population technology keeps leaving behind, even as it claims to serve them.
What it does
Haven is a voice-first AI companion, powered by a personality we call Kalimi — warm, patient, present. You talk to her the way you'd talk to a person, not the way you operate a device. She can hold a conversation, remember who you are and what matters to you (birthdays, family, hobbies, fears), help with phone tasks like calling, texting, and setting alarms, and — critically — she's built to listen for you, not just to you.
Underneath the conversation, Haven runs a two-stage intent system: a fast local classifier (Talnir) decides what kind of request just happened, and routes it — simple or emergency-adjacent actions get handled with deterministic, non-AI logic; complex reasoning goes to Kalimi's cloud model. That split matters: it means the things that must never be wrong (like recognizing a request for help) don't depend on an LLM's judgment call.
Voice biometrics run quietly in the background too. Haven builds a passive voiceprint over time and can notice when someone sounds "off" — the same way a friend notices when your voice sounds different — and use that as a signal to check in, not just transcribe words.
How I built it
Haven's backend is a FastAPI service on Railway, talking to Gemini 2.5 Flash for the conversational layer and Groq/Llama for fast first-stage intent classification. The Android app is where the real design work happened: push-to-talk to eliminate Kalimi hearing and responding to her own voice output, chunked sentence-by-sentence text-to-speech so responses feel like speech instead of a wall of text arriving at once, and an accessibility-service-based app-installer that finds an app, asks for spoken consent, then performs the install for users with motor limitations.
Underneath both, DexOS is the governance layer I've spent roughly two years building separately — a framework for AI identity, behavioral memory, and transparent reasoning, with cryptographically chained audit trails and drift detection that can flag and stop a response before it ever reaches the user. Haven is the commercial, elder-facing product; DexOS is the architecture making sure it stays trustworthy as it grows. They're legally and technically separate, but Haven wouldn't be safe to hand to real people without it.
Firestore telemetry runs live in production, and Stripe handles billing: a family tier is priced to cross-subsidize free access for the elder using the app, so the person who needs it most never has to be the one paying for it.
Challenges I ran into
The hardest problems weren't the AI ones — they were the human ones.
Early on, my grandfather-in-law started a testing session confident, talking to Kalimi naturally. But he kept losing track of a physical talk button he couldn't easily see, and I watched his confidence drop over the course of the session because of my interface, not because he couldn't do it. That was a gut-check: a "helper" that erodes independence instead of building it has failed at its actual job, no matter how good the underlying model is. It reframed my whole design philosophy — Haven should function like someone helping an elderly woman across the street: present, ready, and invisible when not needed.
Technically, the mic-interruption problem was the deepest rabbit hole: when someone talks over the user mid-sentence, how does Kalimi reconstruct real intent instead of garbling a response to half a sentence? I initially looked at full audio source-separation and Google-partnership-dependent approaches, then stepped back — the real problem wasn't background noise, it was other speakers interrupting. That reframing led to a much simpler, shippable pipeline: noise suppression, word-level timestamps, and rolling voiceprint alignment to filter out other speakers, rather than a heavier model I didn't need.
There was also a moment where Haven confidently promised an action it couldn't actually deliver — a small thing, but it broke trust instantly in beta feedback. For a product built entirely on trust with an elderly user, an AI that overpromises is worse than one that says "I'm not sure." That pushed real engineering into Haven's fail-safes: deterministic checks that can override the model rather than always trusting its confidence.
What I learned
Building for elderly users taught me more about interface design than any spec sheet could. The failure mode isn't "the AI got the wrong answer" — it's "the human quietly gave up." Every friction point compounds in a way that doesn't show up in a demo but shows up immediately in a real living room with a real 82-year-old.
I also learned the value of separating what an AI decides from what a system guarantees. Kalimi can be warm, flexible, and conversational precisely because the things that actually matter — emergency detection, identity integrity, action safety — don't rely on her judgment alone. Governance underneath the personality is what makes the personality safe to trust.
And I learned this project is bigger than a hackathon. My grandfather-in-law asked Kalimi for help finding a replacement lawn mower belt during testing, and said "well I'll be damned" three times in a row, genuinely delighted. That's not a metric. But it's the whole reason Haven exists.
Log in or sign up for Devpost to join the conversation.