The financial services industry is undergoing a significant transformation as artificial intelligence (AI) becomes an integral part of its operations. However, the adoption of AI also brings challenges related to transparency, compliance, and risk management. Addressing these critical issues, Arya.ai has introduced AryaXAI, a groundbreaking tool designed to ensure explainability, monitoring, and alignment of mission-critical AI models. Rajeev Ranjan, Editor, Digital Terminal, interacted with Vinay Kumar, Co-Founder of Arya.ai, to explore how AryaXAI is revolutionizing the BFSI segment and setting new standards for AI governance and scalability.
Rajeev: What is AryaXAI, and how can it revolutionize the BFSI segment?
Vinay: AryaXAI is an explainability and alignment tool to ensure any mission-critical AI model is accurately explainable, monitored and safe to scale. For Financial Institution, using blackbox ‘AI’ models doesn’t offer the needed transparency, limiting the usability of these models or sometimes makes them unusable! And as these models becoming important to the core business, any failure or wrong predictions can lead to substantial financial losses and damages trust, making it not only a business continuity risk but can also lead to regulatory challenges as well. Hence, it is important for these mission-critical AI solutions to be explainable, monitored and safe to use. While FSIs realize this, the explainability and alignment is not coping with the speed of the algorithmic innovation making it more harder to solve this. This is becoming a growing risk inside the FSIs.
Not just businesses but regulators across different geographies understand the need for regulating this and are working on regulating ‘AI’ usage in high sensitive use cases in financial services industry. In August 2024, RBI has introduced Model Risk Management guidelines and EU had enforced AI Act in 2024. It reinforces the need for model risk management and auditability of the models, making it mandatory to have these layers to use AI in FSIs.
Arya.ai stands at the forefront of this and have been working on building explainability for more than 3 years. AryaXAI makes it easy for FSIs to understand AI models and provides various tools to align and scale model risk management. This transparency delivers trust among stakeholders, including regulators, customers, and internal teams. By proactively monitoring for data and model drifts, AryaXAI helps maintain consistent performance and fairness in AI-driven processes.
Additionally, AryaXAI's policy control features allow institutions to implement business and ethical guidelines within their AI systems, aligning AI operations with organizational standards and regulatory requirements. This capability is crucial for managing risks associated with AI uncertainty and ensuring compliance in a highly regulated environment.
Rajeev: How is AryaXAI addressing compliance and other requirements since BFSI and affiliated sectors are highly-regulated?
Vinay: As AI usage getting regulated, it is important to ensure these AI models are compliant and auditable. But unlike other compliance requirements which are largely process driven, AI regulations are very technical and complex. For example, there is a need for these AI models to be explainable but there are no standards for explainability and basis on the techniques used to build ‘AI’, often times explainability can be really complex like in LLMs. And for the sake of explainability, FSIs also cannot use any technique and provide incorrect explainability output. In US, it is required for the AI models to not use certain sensitive features like gender, ethnicity etc. But using surrogate explainability techniques FSIs can hide the use of the sensitive features to regulators. And unlike certain industries, the decisions in highly regulated industries like FSIs, Healthcare etc are open for scrutiny at any point of time sometimes even decades! Traceability and auditability at scale is utmost important in such cases.
AryaXAI created a new technique called ‘Backtrace’ one of the most accurate and true-to-model explainability for deep learning models. It also provide multiple opensource (OSS) explainability techniques like SHAPE, LIME, IG, DeepSHAP, SmoothGrad etc making it easy to create a very highly accurate explainability. ‘Explainability evals’ is one of a kind explainability benchmarking for benchmarking multiple explainability techniques. Enterprises can pick the right explainability technique backed by evals! And AryaXAI scales traceability to not just for models but also for each predictions. Users can trace back the model state for any prediction at any point of time! In addition to this, AryaXAI multiple experimental risk management options to stay ahead of the regulations!
Rajeev: Can you explain how AryaXAI can address AI-related issues like data drift and model bias?
Vinay: Model performance is highly influenced by issues like Data drift and model drift. But at the same time, influence of these issues is not easily correlated! So, it is not only important to monitor drifts but also important to use the right monitoring metrics! AryaXAI monitoring stack provide close 7 different drift monitoring metrics like PSI, KL-divergence, Chi-Square test etc and automates the entire model monitoring efforts using ‘Monitors’. Users can create these monitors and define frequency. All these monitors are easily scalable to any kind of volume! These ‘monitors’ alerts the right team if any drift is observed.
While model bias can lead to poor model performance, certain biases like gender bias, ethnicity bias can cause more damage – regulatory questions, reputation damage and loss of business. AryaXAI provide multiple tools monitor such bias and create ‘monitors’ to monitor such behaviours in production.
Rajeev: What is ML Observability integrated in the AryaXAI?
Vinay: More than observability our focus is primarily around creating the most accurate and diverse explainability coupled with risk management for mission-critical AI use cases. True scale of ‘AI’ can only be realised only when we solve these issues for mission-critical use case! The components in AryaXAI are Explain, Monitor, Align and Risk.
Rajeev: How easy is it to deploy AryaXAI in existing infrastructure?
Vinay: AryaXAI is the most critical value-add to ML stack for enterprises. It can work with any MlOps platforms as we only need the models and the data which can be integrated as real-time or batch functions in any environment.
For AI Researchers and Practitioners, we have python SDK that can be easily in any environment and we have a fantastic GUI for any stakeholder to see all the reports, configure and use the product. The SDK is extremely scalable and we made to super easy for developers to integrate within the notebook and perform all actions of the product right from the SDK.
Rajeev: What role does synthetic data generation play in improving AI model performance and alignment?
Vinay: Synthetic Alignment is as good as the quality of the synthetic models! From LLMs to traditional tabular datasets, it is important to ensure these synthetic models are generate accurate and safe data. If done right, the synthetic data can be used as a very scalable strategy to align models very quickly and accurately. In addition to this, synthetic models can provide a new opportunity to test and validate models proactively. But these models can also become the source of model leakages too! So, hence it is important to test these models on data privacy. In AryaXAI, we plugged GPT-2 and CTAGAN models which are highly accurate for generate tabular synthetic data alongside with robust Data Privacy checks around these data generation. We are planning to add Image & Text synthetic models as well.
Rajeev: How does AryaXAI’s compatibility with diverse infrastructures make it a flexible solution for businesses?
Vinay: We always wanted to use a fully serverless AI notebook! It seemed very rudimentary to use the highest compute option only because 1 task the most compute. While we are few months away to launch our notebook but we enabled fully serverless functions to all the tasks in AryaXAI. But we also enabled dedicated instance options as well to provide flexibility! And all of these are multi-cloud with no commitment to any single cloud. No more tie up! And for use cases where the data privacy paramount, we have full on-premise option that can run the entire AryaXAI in any cloud or datacenter.
Rajeev: What long-term benefits does AryaXAI offer for enterprises looking to scale AI responsibly and effectively?
Vinay: Transparency, Standardization, Safety!
Let the your data science team focus on building, our tools can make sure these explainable, monitored, aligned and safe to scale across the organization, any mode, any use case and any user! For your data science teams, they don’t need to spend weeks and months to continuously try to convince the business user and don’t need to spend days to create these dashboards. For risk manager, they now have a tool that they can use to monitor, and manage risk of these models in production. For business users, ‘AI’ is less of a mystery and explainability can provide much needed transparency. For compliance teams, they can now deliver the regulatory artifacts at any time, any place! For Data Science heads, this is a critical tool to make sure AI deployed responsibly and safely at scale.
𝐒𝐭𝐚𝐲 𝐢𝐧𝐟𝐨𝐫𝐦𝐞𝐝 𝐰𝐢𝐭𝐡 𝐨𝐮𝐫 𝐥𝐚𝐭𝐞𝐬𝐭 𝐮𝐩𝐝𝐚𝐭𝐞𝐬 𝐛𝐲 𝐣𝐨𝐢𝐧𝐢𝐧𝐠 𝐭𝐡𝐞 WhatsApp Channel now! 👈📲
𝑭𝒐𝒍𝒍𝒐𝒘 𝑶𝒖𝒓 𝑺𝒐𝒄𝒊𝒂𝒍 𝑴𝒆𝒅𝒊𝒂 𝑷𝒂𝒈𝒆𝐬 👉 Facebook, LinkedIn, Twitter, Instagram
