Cue

• 

  Cue WEBUI, what you see on the glasses

Inspiration

At a career fair you meet thirty people in two hours and forget all of them by dinner. We're two UCLA freshmen, and we freeze on the opener every time. Every tool for this either makes you look down at a phone or whispers in your ear while someone is talking to you. We wanted the opposite: help in the five seconds before you open your mouth, nothing during, and the follow-up actually sent after.

What it does

Hold the temple of your Even Realities G2 and say "Tim Cook, Apple, chairman." Three seconds later the lens shows three cards: how to open, who they are, what to ask for given your goal and your own projects. Then it goes quiet. One dim line while you talk. If it can't find anything public about the person, it says so instead of making something up.

When you're done, hold the temple again and say what happened. Your phone gets a three-bubble recap by iMessage. In the background a CREAO agent researches what came up and writes the email, next steps, three times to meet, and a tentative hold on your calendar. Reply "email" or "book" when you want them. Nothing is ever sent to the person you met.

Text the line "I'm going to LA Hacks tomorrow" the night before and a Research agent builds a pack of who's there and who's hiring. Ask "how did I do today?" and a Coach agent that has read every conversation, with memory of earlier days, tells you who to chase first and one thing to do differently. Two more scheduled agents read the same journal: a Speech Coach that scores your verbatim debriefs for filler words and pacing, and a Relationships agent that flags stale follow-ups each week.

How we built it

A TypeScript app on the Even Hub SDK, running inside the Even phone app, that holds no keys and talks to one backend. The backend (Node, on AWS behind HTTPS) does the three-second path: Deepgram for speech with the event's names as boost terms, Exa for facts, one OpenAI Decisions API call to decide whether this is really the person you named and whether it's worth your eyes, one Claude Haiku call to write the cards.

Everything slower than three seconds is a CREAO agent, started through CREAO's per-agent API: Research before the conversation, Follow-up after it (Gmail draft and Google Calendar connectors), Coach over the day (Google Sheets, extended memory, nightly schedule). Every stage lands in a journal the agents read on their own, so long-horizon agents don't need our backend at all. Photon's iMessage line carries everything to the wearer.

Agents (IDs and links): Research 5283bd7f-cc11-4b29-902b-2debe558b46f, Follow-up 7a639acb-b15a-4d88-b66d-c76831b92ddf, Coach b095ddf1-cccc-42ae-a381-ccc7906b14a4. Install links are in the Try it out section; run share links: Research run: https://agent.creao.ai/share/4bfe8cec5e35f2b116e00947a4049909 Follow-up run: https://agent.creao.ai/share/b23765406116daad02fef43cb845acab Coach run: https://agent.creao.ai/share/255bb913f81f39b0cff9a4103706a61f Speech Coach run: https://agent.creao.ai/share/c2b9ed57edc1cc407c0d387488f02a24 Relationships run: https://agent.creao.ai/share/28e20c6bfe45d020b53b92bc4e5660d5 Long run Analysis: https://agent.creao.ai/install/6fed239d-48e0-455d-b852-7eb72ce7fb0b?version=c20338f2-3eec-4beb-829d-367280c2975f Graphic Builder: https://agent.creao.ai/install/3513fc4c-30a8-4001-af22-155050bb9c65?version=4a594485-5ab7-4995-a7a2-692d0a00beb8 Pull the CREAO keys and Cue is three cards and a recap: no event pack, no email, no times or hold, no coaching.

Challenges we ran into

• The lens cut lines in half until we found the firmware's fixed 27 px line height and rebuilt the layout as exact multiples of it.
• Deepgram can mishear names, so we boost terms, use name-and-org filtering, and snap to the attendee list within two letters.
• An agent started over the API can't touch your calendar unless its autonomy is set to Auto. It silently skips instead; this took us an hour to find.
• The first agent texts read like consultant reports. One page of voice rules and a hard cap of three bubbles per text fixed it.
• The UCLA wifi blocks phone-to-laptop traffic, so we use an AWS box.

Accomplishments that we're proud of

• It runs on real glasses. Spoken lookup, debrief, coach, goal change, and consented recording, all on the G2 against the deployed backend.
• When not enough information is found, it says "no public info found" instead of lying.
• Three CREAO agents doing the heavy work before, after, and across the day, and a journal that lets more agents join without touching our code.
• No camera, no face recognition, no recording without a consent screen the other person sees you tap, and nothing ever sent to them.
• We’re able to convert these conversations to quantitative data and graph it allowing you to track your networking and conversation progress

What we learned

On a 576 by 288 monochrome lens every character is a decision, and measuring on hardware beat reading the docs. "Quiet" is a feature you have to build. Agents belong where waiting is fine; the live path has to answer in seconds. Use a decision model to decide and a generative model to write.

What's next

Past-conversation memory on the Connect card ("you met Priya at the fall fair"), a "send" reply that sends the follow-up from your own account (today you press send yourself, on purpose), and one-tap import of an attendee list.

Built With

• amazon-web-services
• anthropic
• claude
• creao
• even-realities
• photon