At Dignep, we start artificial intelligence development by creating a 100% secure environment for data processing and storage during AI development, applying our best security management system and DevSecOps best practices throughout the SDLC. If we use sensitive data to train an AI model, we anonymize it to avoid the risk of data breaches.

To make sure the AI solution itself doesn’t pose unnecessary risks, we implement data encryption at rest and in transit and robust role-based access control mechanisms. Additionally, we employ data masking, enforce strict logging and monitoring practices, and utilize advanced threat detection mechanisms such as ML-based intrusion detection.

Our compliance experts make sure the AI solution aligns with regional and industry-specific regulations and standards, including HIPAA, GDPR, KYC/AML, and more. We also guarantee transparency for users: they get clear explanations of what personal data is being collected and what for and are asked for consent to data collection and processing.

In such cases, we recommend starting with a proof of concept to check the idea’s feasibility in the shortest possible timeframe. Designing a proof of concept (PoC) is a good way to showcase how the solution will work, estimate the potential value, address major concerns, and draw up a risk mitigation strategy. PoC is also the best choice for a startup company to get a demo version of the future app and use it to attract investments. PoC is highly recommended for innovative AI solutions, where there may be several technology choices that haven't been tested before.

Indeed, data quality largely determines AI output accuracy. However, quality is not an inherent or objective attribute of any data set. Each project has different requirements, so even if the quality of your data is lower than expected, our data engineers can improve it to achieve the desired level. Our professionals use automated tools to assess, cleanse, and deduplicate the data to avoid human error and save time. In case your data is insufficient, we can also enrich it by using external sources (e.g., financial data marketplaces, social media, GIS).

The need for human involvement depends on the case. High-risk tasks like medical image analysis may require constant human presence to verify the AI output, while lower risk tasks (e.g., data entry) will require zero or close to zero human participation. Here are some of the key factors that affect AI output quality, depending on the use case:

• Data quality and quantity. Training data should be clean, relevant to the use case, and representative of the future input that AI will process. Since larger datasets often lead to a higher quality of output, we strive to collect as much data as necessary and can augment the data sets provided by our clients. For example, we can get additional data from relevant online sources with the help of web scraping tools or use generative adversarial networks (GANs) to generate synthetic data for the training set.
• Model selection and training. Depending on the project specifics, we select ML models that will ensure adequate output accuracy and an acceptable cost-to-performance ratio. For highly innovative cases, we develop custom ML models.
• Model validation and testing. We implement robust ML validation and testing mechanisms, including cross-validation.
• Evaluation metrics. We define and apply clear evaluation metrics that align with the AI solution’s goal. Common metrics include precision, recall, F1-score, and mean squared error. We monitor and evaluate model performance using these metrics.
• Human-in-the-Loop (HITL). Depending on the use case and criticality of the output, it may be necessary to implement the Human-in-the-Loop (HITL) system. This involves human reviewers who can validate or adjust the AI output when necessary. It may be recommended for cases like content moderation, medical diagnosing, and legal document review.
• Feedback loop. After every iteration, the AI output is submitted for an expert review. The feedback is then incorporated into the next version of the model to improve its accuracy.
• Monitoring and alerting. We can implement monitoring and alerting systems to detect anomalies or drops in model performance. This allows for proactive intervention when AI accuracy degrades.

Currently, the best way to avoid harmful biases in AI output is to build your software in alignment with UNESCO's Human Rights Approach to AI. To do this, we recommend starting artificial intelligence software development with Human Rights Impact Assessments (HRIA) to identify potential cases where the technology may affect individuals’ rights. When conducting the research, it’s essential to combine domain expertise with feedback from multiple stakeholders, including potential end users and representatives of affected communities.