Architecture

Method

The novel part of this product is not the camera. It is step 3: a private corpus that stays yours, scored against a general model instead of being baked into weights.

  1. 01

    Look

    A still from glasses, phone, or this viewfinder. PoC target: 1 fps at 720p–1080p JPEG. Hardware stays a thin sensor.

  2. 02

    Identify

    A multimodal model names the object, with aliases used as search keys. Discovery is generic — not the moat.

  3. 03

    Retrieve + weigh

    The vault is queried on this server. Top matches, each with a relevance score, go to the model. High-relevance records are treated as primary truth. The model fills gaps. Output carries moat_weight, model_weight, and confidence.

  4. 04

    Speak

    A short spoken line, as if through the glasses, plus the full judgement on this screen.

RAG, not a DSLLM

Fine-tuning would dissolve the moat into model weights — expensive to update, hard to isolate per person. Retrieval keeps documents as an asset. Fine-tune later for tone, not for the facts.

This prototype uses lexical retrieval plus model re-ranking over a small personal file. That is the right first loop: cheap, iterable, and honest about what the vault contributed. Weights, prompts, synonym maps, and the vision/text endpoints are all editable in Admin — including local OpenAI-compatible servers.

Where compute lives

Cloud — SaaS. Phone — B2C, data closer to the wearer. Glasses — same privacy story, different silicon. This PoC is the cloud path by default, but Admin can point identify and infer at a local model so the image never leaves the machine. The vault never leaves the app server except as the top retrieved snippets.

Latency target now: 2–5 seconds from glance to answer. Sub-second is a later hardware and model-routing problem, not a product-shape problem.