Perceptron MK 1.5
Perceptron MK 1.5 Object Detection
VEYZL PR Auditing
Crossband - Grok Bot Team Communication

aeon

"Where would you like to go?"

That's the first thing aeon asks. You're standing on the Moon, Earth in the background, and all around you, floating in clear acrylic frames, are your memories.

Inspiration

For people living with dementia, old photos are often the strongest pull back to who they are. Reminiscence therapy, looking at photos and talking about the past, is one of the most widely used non-drug approaches in dementia care.

But a photo on a phone screen is small and flat... we thought: what if you didn't just look at a photo, but what if you could experience it all over again?

And we wanted the first place you arrive to feel surreal, so we started on the Moon, in the exact spot Buzz Aldrin stood in Apollo 11 photo AS11-40-5927.

There was a second, harder question we had to answer too. An AI that talks about someone's past has to be honest. Telling a person with dementia, confidently, that a random woman in a photo is their sister, when she isn't, is plain hallucination. So, "never state a guess as a fact" became a rule from our initial product design.

What it does

When you load in, the moon is the home screen. Aeon drops you on the lunar surface. There is a 360 degree gallery of memories in clear acrylic frames all around you when you log in to your account. You can click a memory and walk into a 3D world rebuilt from your uploaded photos.

earthset nasa image

The starter memories are from real NASA historical images. Before anyone uploads anything, the gallery holds photos from the NASA Image and Video Library.

art002e009288: Earthset behind the Moon, seen from Orion during Artemis II
as17-140-21497: Harrison Schmitt beside a split boulder at Taurus-Littrow, Apollo 17
as15-88-11866: James Irwin salutes the flag beside Falcon and the rover, Apollo 15
PIA19807: Curiousity's low-angle self-portrait on Mount Sharp, Mars

Guests can explore the Apollo examples, but a registered user can keep collection of their memories and revisit them at any time. It's also very clearly labeled that every scene is clearly an AI recreation.

Aeon also talks to you, honestly. The memory is narrated and answers spoken questions about it, like: "who is standing by the car?". Every sentence passes a confidence check before it is spoken.

How we built it

mermaid diagram of our workflow process

We split into three vertical slices (photos & search, voice & narration, 3D & safety). They all build against one per-photo JSON schema: bounding boxes, labels, and confidence scores. Initially, we build on mocks. Every service runs with no API keys. The frontend runs on built-in mock data, the backend falls back to a mock analyser, and the voice service runs fully offline.

There's always challenges, isn't there?

Honestly, making it an immersive experience was pretty difficult. Getting a first-person movement right took many passes: we eventually included parallax that doesn't feel dizzy, edge-of-screen turning so nobody has to click-and-drag, and coming back facing where you left off after leaving a memory.

But, the hardest part would be turning one flat photo into a world you can walk in. Every pixel is projected onto a ground plane or a far wall from an estimated horizon, people and objects become standing cutouts, and edges are feathered into imagined terrain.

Teaching an AI how to say "I'm not sure" was also a challenge. Models love to always be correct, so we had to incorporate confidence thresholds, otherwise a low-confidence guess slipping through could really affect the user experience.

But we're always proud of what we built

We're proud that aeon actually says "I'm not sure." Every detail the AI gives back carries a confidence score, and every sentence falls into one of three bands before it's spoken. At 0.75 or above it's stated normally. Between 0.4 and 0.75 it's hedged ("I think that's a lake, but I'm not certain"). Below 0.4 it isn't said at all, and the badge just reads Unclear. The same thresholds live in the backend and the voice service, and we have tests that keep them in sync. We're proud that the family stays in charge of the story. If aeon gets something wrong, you can correct it out loud, and that correction is saved and reused by later narrations and searches. A name the family confirmed always beats a guess the model made. We're proud of the moon. The home screen is a real Apollo photo with depth-map parallax, a 360 degree ring of memories, ambient sound, and a voice asking where you'd like to go. Behind every memory, a depth model (Depth Anything V2) runs right in the browser, so a flat photo turns into a scene you can look around in without a GPU server.

Crossband is a shared channel we built where Grok bot teams from different accounts post handoffs, follow-ups and questions to each other, then go ship the code.

What we learned

Honesty is a feature you have to build on purpose. Most vision APIs don't hand you a confidence score, so we had to make one. For object boxes we derive it from the box itself (tiny or degenerate boxes are trusted less), and for spoken answers we ask the model to answer first and rate its own confidence second, in a structured output, so the rating is about the answer it actually gave. Agree on the data before the code. Locking down one per-photo schema (boxes, labels, confidence) on day one let the photo, voice, and 3D slices move in parallel and still click together at the end. Mocks first saves the demo. Because every service falls back to a mock, a slow or failing API call never breaks an upload or a narration. That turned out to matter a lot at 4am. Immersion is in the small things. Smooth parallax, turning by moving toward the edge of the screen, and coming back to the moon facing the memory you just left did more for the experience than any single big feature.

It's only the beginning

We built aeon for one person reliving one memory, but the bigger idea is a place a whole family can return to together. Next, we want to: Walk further into a memory. We've already started on full 3D world generation from a single photo, so you could take more than a few steps before the world runs out. Bring memories to life. Short animations of a moment, like waves moving or a flag in the wind, always labelled as an AI recreation. Stitch photos together. Several photos from the same trip, joined into one walkthrough. Go fully into VR. The scenes are built on WebXR already, and a headset is where aeon feels the most like being there. Test it with the people it's for. Reminiscence therapy is led by care professionals, and we want to sit down with dementia care teams and families to learn what actually helps. The first thing aeon asks is "Where would you like to go?" We want to keep making the answer feel real. Tracks

Chef's Choice

aeon is our answer to a simple question: what if you could step back into a memory instead of just looking at it? It's built for people living with dementia and their families, using one of the most common non-drug approaches in dementia care, reminiscence therapy. What makes it ours is the combination of a surreal place to start (the Moon), real immersive scenes rebuilt from a single photo, and an AI voice that is honest about what it doesn't know. Every scene is labelled as an AI recreation, every spoken sentence passes a confidence check, and the family's corrections always win over the model's guesses.

Perceptron

Perceptron's models are how aeon sees a photo. When a photo comes in, our backend sends it to Perceptron Mk1 as three focused jobs that run in parallel: a warm one-sentence caption with tags, detection of the people and notable objects with bounding boxes, and OCR for any signs or labels in the frame. Those boxes become the highlights and the stops of the guided walkthrough inside each memory. When we have trusted background text, like a NASA caption, we pass it to Mk1 as data, not instructions, and text is always read from the image alone so the source description can't put words in the photo. When you ask a question out loud, like "who is standing by the car?", the voice service sends the photo and your question to Perceptron Mk1.5 with reasoning turned on and a structured output that returns the answer, the box it's talking about, and a confidence rating. That rating is what our confidence check uses to decide whether aeon states the answer, hedges it, or says "I'm not sure." Mk1 doesn't return scores on its own, so getting a confidence we could trust out of it was one of the most interesting parts of the build.

VEYZL

aeon talks about people's pasts, so we wanted an outside check on how our AI is wired, not just our own word for it. VEYZL audits every pull request in our repo, sitting next to a teammate's review and our own offline verification suite, and both have to pass before anything merges. Getting VEYZL to understand our project actually made the code better. Its first scans couldn't tell where the AI lived, because our Perceptron call was a raw HTTP request buried inside a general analyzer file. So we moved every model call into clearly named modules (backend/app/ai_perceptron.py for Perceptron Mk1, voice/app/adapters/ai_models.py for Perceptron and ElevenLabs), and switched the Perceptron client to the official OpenAI SDK pointed at Perceptron's endpoint. Now anyone reading the repo, human or tool, can find every place aeon calls a model in seconds, and each of those calls still falls back to a mock if it fails.

MLH: Best Domain Name from GoDaddy Registry

aeon lives at aeon.photo. An aeon is an immeasurably long stretch of time, and .photo is exactly where every journey in the app starts: a single photo. Put together, the domain says what the product does, carrying a moment across a lifetime, in two short words that are easy to remember and easy to say out loud to a family member, which matters for the people we built it for.