Augmenticon ONE · MAKE · Computer vision
PILOTRunning in a pilot at a customer site

How a vision model is made. From the first question to a model you can validate.

A computer-vision model checks one defined state at your equipment: a cassette fully loaded, a lock closed, a line connected. We build it with you in eight steps. Each step has a clear owner, and every image and every decision is recorded.

Talk to us See the eight steps
  1. Define the scopeTOGETHER
  2. Capture instructionAUGMENTICON
  3. Set up states, take picturesYOU
  4. Upload to the platformYOU
  5. Label and trainAUGMENTICON
  6. Draft model deliveredAUGMENTICON
  7. Test in your processYOU
  8. Verify and hand over for validationAUGMENTICON → YOU

If the draft misses a state: more images, a new dataset version, a new draft.

WHO DOES WHAT TogetherWe decide it jointly AugmenticonWe do the work YouYour team does it, at your equipment
The eight steps

Nothing is trained on images nobody can trace.

The images come from your equipment, they go straight into the platform, and every step after that leaves a record you can show an inspector.

  1. 01Together

    Define the scope

    We agree on exactly what the model must recognise, and what it must not.

    • The process step and the states to detect, for example "vial in position" or "lock closed".
    • The intended use: what the model decides, and what a person still decides.
    • Acceptance criteria, agreed before the first image is taken. They must be at least as strong as the check the model supports or replaces.
    • The variations the model will meet: lighting, devices, consumable lots, and the rare situations that still happen.
    RESULT
    Approved scope, intended use and class list
    DRAFT ANNEX 22
    §3 Intended use · §4 Acceptance criteria
  2. 02Augmenticon

    Receive the capture instruction

    We write a detailed instruction that tells your team which states to set up at the equipment and how to photograph them.

    • Every state to create, including faults and edge cases, not only the good case.
    • Camera position, distance, lighting and background.
    • How many images per state, and how to vary them: device, lighting, operator, lot.
    • What must stay out of the picture: people, patient data, neighbouring workstations.
    RESULT
    Capture instruction, versioned
    DRAFT ANNEX 22
    §5 Test data that covers the full range of real conditions
  3. 03You

    Set up the states and take the pictures

    Your team creates each state at the real equipment and photographs it, following the instruction.

    • Real equipment, real consumables, real working conditions.
    • Fault states are set up on purpose, so the model learns what wrong looks like.
    • Ideally captured with the same devices that will be used in routine work.
    RESULT
    An image set from your own process
  4. 04You

    Upload directly into the platform

    The images go straight into the Training Data Workbench in Augmenticon ONE, at your site. No e-mail, no shared drives, no USB sticks.

    • Every upload records its provenance: capture device, origin, collection, and whether people may appear in the images. The question cannot be skipped.
    • Each image is identified by its content fingerprint. A duplicate is recognised, and every upload is recorded with who made it and when.
    • Nothing is deleted. An image left out of training keeps its history and the reason it was excluded.
    RESULT
    A traceable image library
    DRAFT ANNEX 22
    §5 Data quality · §6 Access control and audit trail
    READ MORE
    Training Data Workbench · System Description, Ch. 7
  5. 05Augmenticon

    Label and train

    We label the images against the approved class list, freeze a dataset version and train the model on it.

    • Only people qualified against the approved class list and labelling instruction may label.
    • Every box records who drew it. Changing a saved box requires a reason.
    • An independent second pass by one named person measures how well two labellers agree.
    • The data is split into training, validation and test, balanced across device, lighting, lot and the other recorded properties. The test set is sealed: anyone who has seen it may not judge the model against it.
    • A second person reviews the evidence and releases the frozen dataset version by electronic signature.
    RESULT
    Released dataset version, identified by its content hash, and a trained model
    DRAFT ANNEX 22
    §5 Verified labels, stratified data · §6 Independent test data
  6. 06Augmenticon

    Receive a draft model

    You receive a draft version of the model for testing. It is clearly marked as a draft and is not used for GMP decisions.

    • A short summary comes with it: which dataset version it was trained on, how it performed, and its known limits.
    RESULT
    Draft model and summary
  7. 07You

    Test it in your process

    Your team uses the draft at the real workstation and tells us where it is right and where it is not.

    • The draft runs in Augmenticon ONE in test mode: add a computer-vision control to any master record and link the draft model to it. Test mode is not used for GMP decisions.
    • Test with fresh situations the model has never seen.
    • Report misses, false alarms and uncertain results.
    • If a state is weak, we go back to steps 3 to 5: more images of that state, a new dataset version, a new draft.
    RESULT
    Your feedback, and a joint decision that the model is ready
  8. 08AugmenticonThen you

    Verification and handover for validation

    When everyone is satisfied, we verify the model against the acceptance criteria from step 1 and hand it over to you for validation.

    • Verification runs on the sealed test set, which played no part in training.
    • The verification report names the dataset version and the model version and describes the sealed test set. For each state it gives the correct results, the misses and the false alarms, with the confusion between states. It states the confidence threshold, how results below it go to a person, and the known limits.
    • The released model is fixed. Any change means a new version and a new verification.
    • You validate the model for your intended use and release it. Our Validation Services can prepare and execute that work with you; the approvals stay yours.
    RESULT
    Verified model and verification report, ready for your validation
    DRAFT ANNEX 22
    §4 Acceptance criteria met · §7–10 Testing, explainability, confidence, operation
    READ MORE
    Validation Services
For your validation team

The detail behind every step

Customer documentation. Sign in with your customer account to open it. No access yet? Write to info@augmenticon.ch.

SYSTEM DESCRIPTIONHow the system is builtArchitecture, data integrity, security CHAPTER 10Roles and responsibilitiesWho does what, ready for your validation plan VALIDATIONValidation ServicesThree service levels, from release evidence to managed validation USER MANUALTraining Data WorkbenchUpload, labelling, dataset versions, release