Klarity Craft functionality

Extract from klarity capacities, once integrate in documentation will be removes or mark as completed

• [ ] <safenai_member> can push and store its artefacts and metrics on KS

• [ ] <safenai_member> can push and store its datasets set on KS

• [ ] <tech_user> can build the required dataset as specified in associated Blueprint

• [ ] <tech_user> can integrate AI component baseline as specified in Blueprint

• [ ] <customer_other> can easily import Blueprint specified data inside Klarity

• [ ] <business_expert> user can provide the intended purpose and ROI elements (metrics) so that the corresponding artefacts can be created from KC.

• [ ] <safenai_member> can locally create and manipulate artefact before delivering configuration that will generate the final ones to the cloud instance of KD

• [ ] <tech_user> can store its models set on KS

• [ ] <tech_user> can instanciate a blueprint to generate configuration and specific code in KC project

  ./klarity folder in UC repository root to control KC and repo content

• [ ] <safenai_member> can locally develop metrics and artefacts for a specific blueprint

• [ ] <tech_user> can locally work and develop a use-case for a specific blueprint

• [ ] <tech_user> can use a CI to build and push artefact from WB/KC to KD

• [ ] <tech_user> can start metric computation from models/dataset/feedback/spec as specified in Blueprint

• [ ] User can launch computation from the workbench over a blueprint instance of use-case

• [ ] User can develop, test and debug version on the cloud environment for a specific blueprint

• [ ] User will rely on dependancy and computation graph integrated management to regenarate elements impacted by any change and evolution

• [ ] <tech_user> can push AI Component simulation metrics and artefact which allow its computatio by WB before being pushed to Karity Dashboard, as required in Blueprint

Data activities

TODO : match with method activities for consistency

• Data augmentation

  • Diffusion / Gan auto enconding models trained to
    • create data with the contextual informations
  • Data reallistique pertubation
  • Classical data augmentation

• Data selection

  • Which data to use for training / validation / test
    • split strategy
      • random with no overlap in sequence
      • use context to reserve several context for test
      • klarity split
        • use auxiliary model to clusterize the data
        • evaluate distance between exemples
        • add these cluster to context, and correlate to existing metadata
        • create n subset that will be used during training / evaluation / test process
        • create a dummy subset for process check

• Data annotation

  • use SAM for pre anotation
  • mixe real generated data

  Which type of anotations - basic : - true / false (or anomaly intensity) - can be real time with operator - optional tags can be proposed

  - precise anotation (expert) : 
       - anomaly tags (type of anomalies), 
       - precise location of anomaly, object, signal portion, ...)
       - optional verbatim, 
       - possible mixture of expert
       
  ==> Base on type of anotation, who is anotation, level of verification (mixture, quality process) a confidence level will be attach to the anotation
  	
  - A specific elment located in the data
    - an object, a signal portion, a general information for the sample 
  

  ==> Compute metrics on user reliability, model reliability, ... possibly multiple anotation for a same sample ==> Propose automatic anotation with a rational / overlay on the screen

  How do we define anomalie - A distance (from normal sample) ==> during anotation, ask for a level of anomalie range fro 0-10 instead of yes / no - Kind of anomalies (might be several type in a same sample, so an anomaly vector where average ~= anomaly intensity