Today's tech blogger: Umesh Mohite, 鶹 AM CTO Lead for Cloud & AI and 鶹 Technology Fellow

Tracing the roots of AI: from Turing to today

The evolution of AI can be traced back to Alan Turing’s 1950 paper “Computing Machinery and Intelligence” and the formalization of the term “Artificial Intelligence” at the 1956 Dartmouth Workshop organized by John McCarthy. Early AI adoptions focused on logic and reasoning, attempting to capture human knowledge in computer programs. However, they faced limitations in computing power and the ability to handle large amounts of data.

From the 1980s onwards, the emergence of Machine Learning (ML) saw its application in various tasks such as marketing and web searches. This was facilitated by growing computational power and the ability to handle larger datasets.

In 2012, Geoffrey Hinton demonstrated how neural networks and deep learning could process vast amounts of data and make accurate predictions, laying the foundation for natural language processing and speech recognition.

Since 2020, the world has been taken by storm with the emergence of generative AI, powered by Large Language Models (LLMs), which can generate text, images, and videos in response to text prompts. Now, Agentic AI is rapidly emerging as a significant advancement, moving beyond generating responses to taking action on behalf of users.

It’s been a long journey so far - let’s pause for a moment and look at where we stand today at 鶹.

Harnessing generative AI: our strategic framework

Due to the newfound power of generative AI coupled with Agentic AI, excitement across the business world and information technology is multi-fold and everyone is eager to employ it at the earliest opportunity. However, this enthusiasm carries the risk of creating similar AI solutions multiple times in different areas. For example, building multiple chatbot interfaces across different areas can lead to inconsistent user interfaces and features, resulting in duplicated capabilities and wasted effort.

Portrait of Umesh Mohite
We’ve adopted an approach and created an enterprise data and AI platform that is used across our business. This strategy leads to consistent implementations, reduced technical debt, and accelerated AI adoption. We call this platform AI Foundations.
Umesh Mohite, 鶹 AM CTO Lead for Cloud & AI and 鶹 Technology Fellow

AI Foundations is comprised of three pillars:

  • Enterprise Data Mesh and Data Platforms
  • Enterprise AI Suite (which includes model development platforms)
  • AI Platforms (which includes a unified chatbot experience for 鶹, AI services, document management, and Agentic AI platforms).

Turning strategy into action: architecture, intelligent autonomy and the power of our AI working groups

Working groups were created for each topic in AI Foundations, with experts identified from across the business to define our tech strategy. I was part of the Agentic AI and document management working groups. Document management is an equally important topic, but in my opinion, not given enough importance in the AI world.

The goal of our Agentic AI working group was to develop a group reference architecture for adopting the Agentic AI ecosystem. This included defining the required standards, governance, security, compliance and monitoring, efficiency, validation, development, and support target model.

Agentic AI reference architecture can be broken down into:

  • Core capabilities that include an orchestration layer that validates inputs, performs planning by discovering required Agents from the Agent Library, with runtime to execute the agent orchestration. Library/Marketplace for Agent and Tools discovery; and a Memory layer to provide short- and long-term memory, including state management for agent orchestration executions. 
  • Shared services concerning guardrails, logging and monitoring, and security.
  • Gateway or integration layer that allows access to the platform.

The process begins with users, systems, or event triggers accessing the orchestration engine via the Gateway. The Gateway handles concerns such as access security and throttling. The orchestration engine then invokes the planner layer, which consists of reasoning and validation components. The validation component includes input validators that ensure the entry is free from potential threats. Next, the reasoning component is invoked, which, based on the input, accesses the memory and LLM to discover relevant agents from the agent library or marketplace. Multiple agents may be required to complete the task. A plan is formulated, and the identified agents are executed accordingly. Each agent requires its own planning to achieve its task. The agents also access the Memory and the Tools library to find the necessary tools for task completion. These tools are specialized executors designed for specific tasks, such as accessing LLM for summarization.

All of this is overseen by the shared services components. Guardrails at various levels - planner, agents, and tools - ensure compliance is adhered to and further reduce risk by incorporating a human in the loop. Logging and monitoring ensure end-to-end traceability is maintained, while security ensures data protection and access restrictions. Once the orchestrator completes all agent and tool executions, it invokes output validators to ensure the outcome meets the quality standards. If not, a refinement cycle can be triggered.

The Agents and Tools library/marketplace is a central enterprise repository where all agents and tools across the organization are registered with extensive standardized metadata. The platform must be reliable and available, ensuring that orchestrators can consistently access and utilize the library without interruptions or downtime.

The library maintains different versions of agents, ensuring that the most appropriate and stable versions are used in production environments. It supports the full lifecycle of agents and tools, including deployment, monitoring, updating, and decommissioning. Integrated with the AI Governance framework, the library ensures that only approved agents with the right level of data confidentiality and zone limitations can be used.

Agentic AI reference architecture

Memory can be organized into three distinct layers, each serving a unique purpose in maintaining context, facilitating learning, and enabling adaptation over time:

  1. Short-term memory retains recent interactions to ensure contextual continuity within a single session. This includes maintaining the execution state of the session, checkpointing information regarding agents’ state and handoffs between agents - such that the execution can be resumed seamlessly in the event of a failure. As a cache, it processes current inputs, tracks ongoing conversations, and uses attention mechanisms to prioritize relevant information. Due to its nature, it has limited capacity, and as new data arrives, older data is overwritten unless there is a need to transfer the data to long-term memory.
  2. Long-term memory stores data for future reference, such as user preferences, past interactions, learned workflows, or domain-specific knowledge. This enables the recognition of recurring patterns, recall of past interactions, and personalization of responses based on accumulated experiences.
  3. Feedback loops help with the self-improvement of the overall agentic system by refining the plan, agent logic, both short-term and long-term memory over time. This is done by incorporating user feedback, such as ratings, or system inputs, like error tracking.

Guardrails are essential for maintaining consistency, quality, security, compliance, safety, ethics, alignment with human and 鶹 values, and efficiency across the AI ecosystem with minimal human oversight. Key aspects of guardrails include:

  • Establishing clear standards and effective governance structures: defining the required standards for governance, security, compliance, monitoring, efficiency, validation, development, and support target models.
  • Ensuring compliance with relevant regulations and policies: continuous monitoring is necessary to avoid legal issues, maintain ethical standards, and ensure that agents operate within acceptable boundaries.
  • Business or organization-specific rules and conditions: these serve as guardrails and may be general (applicable across agents) or specific (agent-specific).
  • Triggers to ensure human oversight is invoked for the most sensitive agentic tasks.

The trajectory is clear - embracing Agentic AI will redefine what it means to be a financial institution, and we’re moving towards a future where banks are no longer just platforms or service providers but are adaptive, proactive, intelligent ecosystems that act in real-time to serve our customer needs and organizational objectives.