Izengard Merlin - Unified Causal AI Machine Learning and ML Ops : Model Management Platform
Izengard integrates a world class ML ops platform into the product working the data fabric / data mesh and providing many capabilities as outlined below.
Traditional Machine Learning methods often have difficulties with correlation. Correlation is not cause and effect. For this Human In The Loop style domain expertise is used. Izengard brings typologies and we iterate a Causal AI model with your Financial Crime/Cyber-Security and other domain experts to come to a model that can be deployed easily, has more explainability and you have confidence in front of a regulator or legal counsel.
No need to use external machine learning tools, the Causal AI model environment is provided via a third party and integrated into Izengard. Model in a united manner across all 3 domains of Cyber-security, IT Risk Management and Financial Crime and all the rich data and history available in Izengard, your business analysts, data scientists and users can deploy, test in sandbox, version control, modify and run high precision models quickly and from a single platform.
Izengard Merlin - Causal AI Machine Learning : Features and Differentiators - please click numbers on left hand side panel to navigate through the features.
Izengard Optimized Causal AI Models and Algorithms
Izengard works with your SMEs and our partner on Causal Models to help data scientists and your domain experts optimize models, not only from suggesting encodings but also understanding how varying hyperparameters, changing input parameters and understanding confounders relates to the output predictive power of the models. Izengard supports a range of Causal AI algorithms, which we will be happy to disclose to you in person.
Create your own Causal AI models
Izengard is flexible in that it allows clients with training to create their own Causal AI models iteratively with our professional services team and support from our Causal AI partner.
Data Quality & Data Preparation (pipelines)
Izengard supports best practices in Data Pipeline orchestration, involving binning, converting qualitative data to binary, encoding, ignoring nulls, transformations and user-defined features, aggregations and various computations. This prepares the data for feature engineering. Izengard uses Apache Beam and Apache Kafka extensively for dataflow pipelining and streaming
Feature Discovery & Engineering
Izengard supports Feature Causal AI Discovery, which uses confounders to find cause and effect, this will booste accuracy and reduce feature count at the same time. This way feature sprawl and feature redundancy which highlight feature degradation and increasing feature costs are managed. Izengard supports a 6 monthly check on models provided by Izengard to clients to do these checks and to optimize them further as part of the support contract.
Model Development : Jupyter Style Notebook and Git Integration
Izengard through it's machine learning technology partners, makes available a Jupyter style notebook to develop machine learning models. In addition to this GIT integration is there so that version control via check in and check out can be managed.
Model CI / CD pipelines (Source : Google Cloud)
Machine learning provides different challenges to traditional Continuous Integration Continuous Delivery approaches to DevOps. This first issue is that machine learning is iterative and experimental, the second, so track and maintain reproduceable approaches is challenging. Then there are challenges in testing and validation before deployment pipelines and production level training must be considered. Izengard's technology partner for machine learning and Izengard data scientists have best practices deployed and these are further taught to you via our professional services teams and partners.
Model Performance and Drift
Izengard supports different model performance metrics for Causal AI models to monitor the real-time performance of production models. In addition, it measures drift in several ways, including : concept drift, prediction drift, label drift, feature drift. Izengard has a complete approach in it's model management suite to report on drift metrics and to identify causes of drift such that the model can later be replaced or optimized.
Izengard gives you complete control over your hyper parameters and these can be tuned or optimized using our model management optimizer provided and integrated from our technology partner
Model Testing Sandboxes (Model Validation)
For data scientists as well as internal audit, separate entire sandbox environments can be set up to test prior to release in production or to carry out model validation exercises. Whilst the size of sandbox must be determined as it has a cost element towards it, this flexibility can be pre-configured for a certain amount of historical data etc.
Lookbacks - Keeping models which were used in force at the period of the lookback.
In AML as in many crime related scenarios, the detection is often done after the fact. Therefore it is important to often piece together events of the time and the actions taken at that particular snapshot of time. During that time machine learning models may have changed and therefore Izengard keeps a history of when models are created and retired and never deletes a model, allowing a sandbox to be created to conduct a lookback. Causal AI based models have a deeper meaning for lookbacks and therefore current models or models used previously can be used with any timeline under investigation.
Types of Machine Learning Supported
Izengard supports a wide range of Causal AI machine learning algorithms and libraries via our data science technology partner these focus heavily on re-inforcement and ensemble methods.
Rules and Scenario Building
Izengard allows clients to combine rules, statistical models and Causal AI machine learning into specific scenarios that may or may not be typology related. This allows the convenience of seeing if certain sequences are followed and detected and to direct the right actions once detected by close integration to the Izengard Command & Control, Case Management system.
Izengard provides a range of explainability approaches which are inherently easier to understand in Causal AI models. The outputs can be further transcribed in business language to make the output decisions of a model easier to understand.