Skip to content

Technology · Agentic AI

Read the signals of a space. Propose what comes next.

It observes and interprets a space, in place of people or alongside them.

Every proposal carries its evidence.

The core argument

Before AI can work in a space, the space has to become data.

Documents and transactions are already data, so a language model can sit on top of them today. Stores and sites are different. A space has to become data first (DX) before AI can work on top of it (AX).

Computer vision based spatial AI is what creates that DX. Adopt the two separately and the work of connecting them remains.

  1. Anonymization

    Behavior signals remain after identity is removed. This is where data you are allowed to keep begins.

    More on Anonymizer
  2. Spatial coordinates

    Flow, dwell, and shelf contact land on one coordinate system. Here a space becomes data you can query.

    More on Spatial AI
  3. Operating standards

    Different stores are counted by the same definitions. An answer from one store holds in another.

Model composition

We do not hand everything to one model.

The vision models that read a space and the language models that put it into words are good at different things. We pick and connect the right model for each task.

  1. Our own vision models

    We build the part that reads flow and state without identifying anyone. Video is anonymized on site before anything else touches it.

  2. Foundation models, combined

    Summaries and dialogue run on published language models. We do not rebuild them. We use what they already do well.

  3. Grounding in spatial context

    The language model receives anonymized spatial data together with the operating standards. It answers about this space instead of in general.

Which combination works better is checked against the record of approvals and rejections. That record is also the reason to swap a model.

How a proposal is built

Behind one proposal there are four steps.

Detect the signal, analyze the evidence, propose the action, and let a person confirm. Pick a case and follow all four steps.

Pick a case
Example data
  1. 01

    Detect

    saai care

    • 7 people waiting at checkout
    • Average wait 4 min 12 sec
    • 1 lane open
  2. 02

    Analyze

    saai insight

    • The average wait for this weekday and hour is 2 min 10 sec.
    • Entrance visits are close to usual. What grew is dwell at the checkout.
    • Over the past 4 weeks the wait passed 4 minutes on 3 days, all of them rainy.
  3. 03

    Act

    saai agent

    Open one more checkout lane.

    Priority · High

    Why this proposal

    The 4-week record shows more people leaving without paying once the wait passes 4 minutes.

    Expected · Wait around 2 min · example estimate

  4. 04

    Confirm

    A person

    From here on, it belongs to a person.

Principles

We drew the lines it will not cross, first.

However capable the agent becomes, these three lines stay.

  1. The AI stops at the proposal

    Nothing runs without approval. Even after approval, it can be stopped and rolled back.

  2. No claim without evidence

    What the proposal is based on sits right next to the proposal.

  3. A rejection stays on record

    When a person says no, that judgement is kept and shapes the next proposal.

How these three lines held up in the field is recorded in a case study.

A hypothesis, a shelf change, and a second measurement

Feature standard

Before a feature ships, it must pass five checks.

This is how principles become product criteria.

Check 03 · Does it reduce what people need to read?

An example placing a legacy alert stream beside a prioritized view. It narrows today’s work instead of adding metrics.

Explore adoption

Decide where to start together.

Tell us how you operate today. We will propose a level to review.