How to Develop AI Software

Duration: 1 month

• Outlining high-level software requirements (in case of new software).
• Creating a proof of concept (PoC) for AI to check the technical and economic feasibility of enriching software with it, estimate the scope of work, timeline, budget, and risks.
• Calculating ballpark ROI of AI implementation.

Duration: 1-6 weeks

ScienceSoft defines detailed functional and non-functional requirements to AI, such as the required level of AI accuracy (in some cases, the value can be driven already with 65-80% of accuracy), explainability, fairness, privacy, and the required response time.

ScienceSoft's best practice: When choosing a machine learning model AI will leverage, we carefully consider the trade-offs between requirements to AI (as, for example, some models can be less accurate but more explainable and fair).

Duration depends on the overall complexity of software functionality

ScienceSoft selects integration patterns and procedures. We design the architecture of the solution with integration points between its modules, including integration with an AI module.

Duration: 1-3 months

Launching an initiative of integrating AI in business-critical software may require organizational changes to increase the chances for its successful implementation and adoption. ScienceSoft discusses with the business:

• Shifts in data policies to break down data silos across the departments to enable easy access to data and avoid duplicated or contradicting data that decreases AI accuracy.
• Determining a plan on adapting employees’ workflows to the use of updated (or new) software (e.g., user training and refreshed user guides and policies).
• Promoting continuous collaboration between business and tech stakeholders.

Duration: 3-36 months

ScienceSoft develops the front end and the back end of software (the server side and APIs, including necessary APIs for AI module integration). We also run all necessary QA procedures throughout the development process to validate software quality.

1. Data preparation

Duration: 1-2 weeks (this process can be reiterated to increase the quality of AI deliverables)

• Consolidating data from relevant data sources (internal and external, which can be acquired via one-time purchase or a subscription).
• Performing exploratory analysis on data to discover useful patterns in it, detect obvious errors, outliers, anomalies, etc.
• Cleansing data: standardizing, replacing missing or deviating variables, removing duplicates, and anonymizing sensitive data.
• The resulting data is split into training, validation and test sets.

ScienceSoft's best practice: To significantly streamline this time-consuming stage, we use automation tools (e.g., Trifacta, OpenRefine, DataMatch Enterprise, as well tools within leading AI cloud platforms – Amazon SageMaker, Azure Machine Learning, Google AI Platform).

2. ML model training

Duration: 1-4 weeks (depending on the model’s complexity)

ScienceSoft selects fitting machine learning algorithms and builds ML models. We trained the models with training data and test against a validation dataset, then we increase their performance  by fine-tuning hyperparameters. The most high-performing models can be combined into a single model to decrease the error rate of separate models. The final ML model is validated against a test dataset in the pre-production environment.

Duration: 2-4 weeks

The configuration of the AI deployment infrastructure and approach to integrating AI into software depends on how AI should output results:

• In batches: AI outputs are cached according to pre-scheduled time intervals. Targeted software retrieves AI outputs from the data storage it is connected with. Higher latency is acceptable.
• As a web service: near-real-time outputs triggered by a user or a system request via API. Low latency is required.

At ScienceSoft, we often start with pilot deployments to a limited number of software users to verify the smoothness of AI integration with target software and compatibility with the infrastructure (latency, CPU and RAM usage) and run user acceptance tests to handle possible issues before a full-scale rollout.

ScienceSoft's best practice: To accelerate the AI deployment, in our projects we leverage leading AI cloud platforms – Amazon SageMaker, Azure Machine Learning, Google AI Platform.

ScienceSoft tracks and fixes software bugs and issues of integration with AI, optimizes software performance and enhances UI based on user feedback, develops new features and extends AI-enabled functionality drawing on evolving business or user needs.

Maintenance of AI is a separately controlled process. It includes monitoring of ML model performance to detect a ‘drift’ (decreasing accuracy and increasing bias when the data that AI processes grows and starts deviating from the initial training data).

In case of the drift, models should be retrained with new hyperparameters or newly engineered features reflecting shifts in data patterns. They can also be replaced by challenger models with higher performance (identified during A/B testing).