Guide — cyber:cyber

// services

Spatial

Agentic

On-device

// Live website

https://www.cybercyber.ai/journal/building-before-the-door-opens

Guide is a hands-free agent for smart glasses: ask it to read a sign, describe the room, or name the thing in your hand, and it answers out loud. We started building it before the platform was ready, because a studio that only builds on finished platforms is always late. Every answer is produced on the phone, by models that ship with the OS — no servers, no uploads, nothing about where you are or what you're looking at leaving your hands. It began as an accessibility tool, which is the version of this problem with no room to be vague.

Building before the door opens

The obvious time to build for a new device is once the platform is finished. We started earlier, deliberately. Guide was built against a toolkit still at version 0.8, on hardware we didn't have yet — and the first useful thing the project produced wasn't code, it was a finding: the two developer paths sold under one brand name are not interchangeable, and the one everybody assumes you'd use cannot reach a camera or a microphone at all. An hour of reading killed a plan that would otherwise have died in week three. Verify the capability you actually need, never the platform it's sold under.

Two models doing a third one's job

Guide has to look at something and then talk about it, and on today's iPhone no single on-device model does both: the vision models don't speak, and the language model can't see. The easy answer was to ship a half-gigabyte download to close the gap. Instead we split the job — Apple's Vision framework does the seeing, the on-device language model does the saying. That turned out to be the better architecture for one specific reason: the language model never receives the image, so it cannot invent an object that was never there. It can only phrase a fact it was handed. For someone who can't check the answer themselves, a smaller failure surface is worth more than a larger vocabulary.

Cheap first, and never go quiet

Guide reads a sign in about forty milliseconds and spends two seconds only when the question genuinely needs two seconds — it starts with the cheapest capable model and climbs only on failure. Getting that right meant learning that confidence isn't a single currency: three models reporting 0.64 do not mean the same thing, and early on the system hedged a sign it had read perfectly, character for character. Miscalibrated confidence doesn't just misinform, it teaches you to distrust the one part that's almost never wrong. Underneath it all sits a single rule — silence is forbidden. When a model failed outright in testing, the system climbed, hedged, and still spoke. What Guide is not is obstacle awareness, and it will never be sold as such. The largest gap is still open: no blind wearer has used this yet, so every claim here about what someone needs is a hypothesis we intend to test, not a result we're selling.