Skip to content
The room that serves what the group runs on
  1. Conglomerado
  2. Calaa Lab
  3. Development
  4. Calaa

01.1

The group's neural network, and who sustains it.

Of the three developments, this is the only one the division does not do alone

Calaa is the group's intelligence layer and runs beneath its three products. It is also a company of the group, so the split of work does not resemble that of the other two developments. Calaa Lab helps develop its neural network and maintains the infrastructure on which it is trained and from which it answers.

01.1.1

The development lasts as long as the products do

Of the three developments Calaa Lab takes part in, this is the only one that does not end with a published version. The application ships and is corrected, and the device leaves with its board. The network keeps learning while the other two are in service, so the work lasts as long as the products. The two decisions that weigh most on this division come from there. It is built inside the group, and it runs on the group's machines.

Where each development ends

Two of the three reach a published version

The three lanes order, they do not measure. There is no time scale underneath.

Cal-1
the version that ships
Cal-1 Link
the board that ships
Calaa
with no point at which it closes

development

  • Cal-1
  • Cal-1 Link
  • Calaa

in service

  • the version that ships
  • the board that ships
  • with no point at which it closes

Select a lane to see where it ends.

developmentin servicewhat runs along the lane

A development with no delivery date is not planned as a project. It is sustained as a capability, and that is the reason this division owns its machines instead of renting them by the hour.

01.1.2

Four capabilities a general model does not bring

A general model arrives knowing about the world and knowing nothing about this sector. They are built one by one, and they are the engineering this division contributes.

  1. 01

    Understanding the question

    A noise, a warning light, a smell

    Nobody describes a fault by the name it has in the catalogue. What gets described is a noise when braking, a light that comes on when cold, or a smell. Getting from there to a specific part number is the first capability that has to be built.

  2. 02

    Speaking the vocabulary of the sector

    Three catalogues for the same part

    The same part is named one way in the manufacturer's catalogue, another in the distributor's and a third at a workshop counter. Recognising all three as the same part is what prevents buying it twice.

  3. 03

    Reading a vehicle's series

    What arrives from the Cal-1 Link

    A single reading says little. The same measurement across months separates normal use from what is beginning to drift, and that reading over time is built apart from the one that answers a query.

  4. 04

    Telling one unit from a series

    A repair and a pattern

    One vehicle with a problem is a repair. The same problem repeated across many vehicles of the same year is a pattern, and telling them apart does not come out of the model on its own.

The two paths of an answer

The four split into two paths that end in the same place

Where each contribution enters and where they meet. The boxes are the four from the list above.

what someone says
Understanding the question
Speaking the vocabulary of the sector
what the vehicle sends
Reading a vehicle's series
Telling one unit from a series
an answer

what someone says

  • Understanding the question
  • Speaking the vocabulary of the sector

what the vehicle sends

  • Reading a vehicle's series
  • Telling one unit from a series

Select a piece to see which path it belongs to.

the four capabilitieswhat enters along each path

All four have to be in place for any one of them to be of use. A question that is well understood but ends in a part number that exists in no catalogue gives a wrong answer with every appearance of being right.

01.1.3

What comes out of the product goes back into the layer

The material the layer is trained on is not produced by this division. Calaa Mind measures it and checks it before it enters, and Calaa decides what the layer has to come to know. Calaa Lab supplies the engineering and the machines, trains, and answers for the result staying in service. None of the three does the work of the others, and that is why it can be asked separately where an answer came from and who computed it.

What turns that split into a circuit is the return leg. The application and the device are used every day, and that use produces new material which enters through the same place the first material did. Each turn leaves the layer knowing something it did not know before, and the two products answer with it without anyone touching a line of their code.

The dependency between the two products comes from there. They hang from the same trunk, so a stalled leg reaches both at once. It does not show on the first day. It shows when the layer stops learning and the answers age at the same rate in the application and in the device.

The circuit that feeds the layer

The return leg is what makes the layer improve on its own

The five stages of the circuit and the two products that hang from it. The drawing orders them and measures none.

the vehicle parc, on the road
Calaa Mind
Calaa
Calaa Lab
this division
the network, in service
Cal-1
Cal-1 Link
what happens in use goes back in

The circuit that feeds the layer

  • the vehicle parc, on the road
  • Calaa Mind
  • Calaa
  • Calaa Lab
  • the network, in service

a single trunk for both

  • Cal-1
  • Cal-1 Link
  • what happens in use goes back in

Select a piece of the circuit to see what it does.

what this division sustainswhat is in the circuit and is not its ownthe flow, and its return leg

The two products are not fed each on its own side. They hang from the same trunk, so whatever happens to any leg above reaches both at once. That is the advantage of having a circuit, and also what obliges it never to stop.

01.1.4

What has to be exact does not go through the network

Some questions have an answer that admits verification. An equivalence between two part numbers, a torque setting, a service interval. Those the network does not answer. They come from a table and are checked against it before being returned. Approximating a figure that exists exactly makes the product worse and makes it look better.

The remaining limits are not decided in this division. The scope of what enters the corpus is set by Calaa Mind. What the layer knows how to answer and where it declines to do so is published by Calaa. Calaa Lab answers for the network being trainable, for it staying in service, and for what is exact never passing through it.

What the network answers and what it does not

Two of the four are not answered by the network

Four queries from the sector and where each one comes from. All four are asked with the same words.

What is askedWhere the answer comes fromWhat makes it good
What may be happening, told in the words of whoever is askingThe networkThat it ends in a part number that exists
Which manufacturer part number matches a distributor'sA tableThat it matches the catalogue of origin
What torque a specific part is tightened toA tableThat it matches the assembly sheet
Whether a measurement on the vehicle is driftingThe networkThat what was measured earlier on that same vehicle supports it

The path is decided by whether a source exists to check the answer against. A hard question with a source comes from the table, and an easy one without a source comes from the network. When the source exists, the network does not come in, and that is an engineering decision of this division.