A companion that remembers, for people who are starting to forget.

The existing tools split into two camps, and both fail the same person. Alarm and reminder apps fire a notification, get dismissed, and learn nothing. Monitoring systems watch for falls and emergencies but treat the person like a sensor reading. Neither one talks. Neither one listens. Neither one remembers that Mary takes the yellow pill, not the white one, and that she calls it “the heart one.”

Caregivers, usually family, get the worst of it. They either hover with constant phone calls that erode their parent's independence, or they stay back and live with the anxiety of not knowing. There was no middle ground: informed without intrusive.

That was the brief we gave ourselves at CalHacks 12.0: not a louder alarm, but a companion. Something an elder could talk to, that learns who they are, and that quietly keeps the family in the loop.

On the elder's side, everything happens by voice. Elda delivers reminders conversationally, asks whether the medication was actually taken, and nudges again in five minutes if it wasn't. Between reminders it checks in every couple of hours, watches for unusual inactivity, and keeps a persistent emergency button one tap away. At midnight it writes the caregiver a daily summary.

The intelligence is three layers working together. Claude handles the real-time understanding: what the person said, what they meant, how to respond with warmth instead of canned phrases. Letta gives Elda long-term memory, so it learns routines and preferences across days instead of starting cold every conversation. Chromaadds semantic search over everything said before, so “the heart pill” finds the right medication even though no one ever typed those words into a database.

Click to view full size

Even onboarding respects who this is for. An elder never creates an account or types a password. The caregiver sets everything up on the dashboard and the elder joins by scanning a QR code or entering a six-digit code, once.

Click to view full size

The stack underneath: a FastAPI backend with schedulers for check-ins, activity monitoring, and alerts; a React Native mobile app with live transcription and text-to-speech; and a Next.js caregiver dashboard.

The core trust problem of any voice interface: silence. The moment a user speaks and nothing visibly happens, they stop trusting the product. So the interaction model came first, four explicit voice states, idle, listening, processing, and speaking, each with its own visible feedback, because an elder with memory loss should never have to wonder whether they were heard. Captions accompany every word the app speaks, so a hard of hearing user never loses the thread.

The same discipline set hard accessibility floors: tap targets at least 52dp, body type at least 16pt, one primary action per screen, a persistent emergency path that is never more than one tap away.

And the microcopy, which carried more of the product than any screen. Elda never says “Task incomplete.” It says “Did you take it?” and then “Great job, I marked it done.” In a product with no visual interface to lean on, the words and their timing are the craft.

With a weekend on the clock, I didn't draw every screen and hand off mockups. I wrote the design system and screen specifications as briefs precise enough for AI to implement directly: the palette, Playfair Display and Nunito Sans, an 8-point grid, motion under 300 milliseconds, every state and every line of copy. That trade is exactly why the visual layer stayed rough and the interaction model got the attention.

A reminder that says “It's time for your yellow pill on the table” is design. The color of the pill and where it sits is the interface.

The caregiver dashboard got its own spec with the same system wearing a more operational face: patient cards, schedule configuration, alert queues, daily reports. Two audiences, one visual language, so the product feels like one thing built with one set of hands.

What I think landed with the judges is that Claude wasn't bolted on for a demo moment. It sat at the center of a system with real architecture around it: memory that persists, search that understands meaning, schedulers that act without being asked. AI used as care infrastructure, not as a chatbot with a nice wrapper.

For me, Elda is also the clearest proof of a way of working I keep returning to. The same week-long rhythms I use on client work, understand the person, write the system down, let AI carry the implementation, held up at hackathon compression. Design didn't slow the team down. It was the reason the demo felt like a product.

The team. Elda was built in one weekend at CalHacks 12.0, October 2025, by a team of three. Gaurav Chaulagain led the engineering across the backend, mobile app, and dashboard, Medhavee Upadhyaya completed the team, and I owned the product design: both surfaces, the design system, the voice states, and the microcopy.