Applied AI · Computer vision

Vision models measured before they are used.

We design and train models that classify, inspect or count from images. We test them on real data and define which cases a person reviews.

01Task and decision defined
02Data and limitations reviewed
03A clear next technical step

What we have learned in the field

Average accuracy hides the errors that matter.

A model can be right almost every time and still fail in late-afternoon light or on a rare defect. So we measure it by type of case, not just on average.

01

Computer vision

Classifying, detecting or counting in production, lab or field images. Sometimes the biggest improvement comes from how the photo is taken.

02

Model evaluation

We test on images the model has never seen and review its errors one by one. That shows where to trust it and where not to.

03

Integration into operations

In the field, images change: another camera, other lighting, a new supplier. We set up monitoring to notice when the model starts losing accuracy.

How a model reads an image

From pixels to a prediction.

A conceptual journey: inspect the input, extract signals, process them in a network and compare the output probabilities.

01

Input

Pixel matrix · 3 × 3 window

02

Features

Signals the model can process

03

Neural network

Propagation through connected layers

04

Prediction

A 92.4% B 5.1% C 2.5%

Output distribution · softmax

Simplified schematic using simulated data. Probabilities illustrate model output, not a guarantee of accuracy.

How we move forward

Test before going live.

Each stage ends with a result that decides whether we continue.

  1. 01

    Feasibility

    Available data, the task and the success criterion.

  2. 02

    Focused test

    An initial model compared with the agreed criterion.

  3. 03

    Operation

    Integration, human review and performance monitoring.

A practical question

What happens when the model gets a case wrong?

An AI solution needs a way to evaluate errors and decide how to respond. With your team, we define what can be automated, what needs review and what evidence will support the next step.

A DECISION WE CAN REVIEW Estimate Criterion metBASED ON EVALUATION Human reviewCASE TO EXAMINE
Conceptual scene · no client data or results

What your team can review

  1. 01

    Feasibility and evaluation criteria

    A focused task, a review of available data and an agreed way to examine results.

  2. 02

    Models and evidence of limitations

    When development is in scope, we work with models, evaluation results and error examples the team can review.

  3. 03

    An integration path

    Defined inputs, outputs, human review and conditions for using and maintaining the solution.

When it makes sense

When someone reviews images over and over.

And there are enough examples to teach a model what to look for.

  • 01

    Repetitive visual inspection in production or the lab

  • 02

    Accumulated images nobody has time to review

  • 03

    Classification criteria that vary by who is looking

  • 04

    An AI pilot that worked in testing but not in operation

Before we talk

A useful starting point.

You do not need to arrive with a model selected. Start with the task and some representative examples.

01

Which task repeats?

Describe what a person looks at or classifies and what they do with the result.

02

What examples exist?

Identify images or data, their source and who understands the categories. We can start with a description of the sources.

03

Which errors matter?

Explain what happens if a case is missed or an incorrect alert is raised, and who could review it.

A related caseUniversidad Santo Tomás · ANID projectEngineering, modelling, training and software for fresh-smear microscopy image classification.

Let's talk

Tell us what you need to solve.

We give you a technical read of the problem and a possible starting point.

Let’s talk