Harnessing the ability of generative AI: Druva’s multi-agent copilot for streamlined knowledge safety

This publish is co-written with David Gildea and Tom Nijs from Druva.

Generative AI is reworking the way in which companies work together with their clients and revolutionizing conversational interfaces for advanced IT operations. Druva, a number one supplier of knowledge safety options, is on the forefront of this transformation. In collaboration with Amazon Internet Companies (AWS), Druva is growing a cutting-edge generative AI-powered multi-agent copilot that goals to redefine the shopper expertise in knowledge safety and cyber resilience.

Powered by Amazon Bedrock and utilizing superior giant language fashions (LLMs), this revolutionary resolution will present Druva’s clients with an intuitive, conversational interface to entry knowledge administration, safety insights, and operational assist throughout their product suite. By harnessing the ability of generative AI and agentic AI, Druva goals to streamline operations, improve buyer satisfaction, and improve the general worth proposition of its knowledge safety and cyber resilience options.

On this publish, we study the technical structure behind this AI-powered copilot, exploring the way it processes pure language queries, maintains context throughout advanced workflows, and delivers safe, correct responses to streamline knowledge safety operations.

Challenges and alternatives

Druva needs to successfully serve enterprises transferring past conventional query-based AI and into agentic programs and meet their advanced knowledge administration and safety wants with larger velocity, simplicity, and confidence.

Complete knowledge safety necessitates monitoring a excessive quantity of knowledge and metrics to determine potential cyber threats. As threats evolve, it may be troublesome for purchasers to remain abreast of latest knowledge anomalies to hunt for inside their group’s knowledge, however lacking any menace indicators can result in unauthorized entry to delicate info. For instance, a world monetary companies firm managing greater than 500 servers throughout a number of areas at the moment spends hours manually checking logs throughout dozens of programs when backup fails. With an AI-powered copilot, they might merely ask, “Why did my backups fail final evening?” and immediately obtain an evaluation exhibiting {that a} particular coverage replace precipitated conflicts of their European knowledge facilities, together with a step-by-step remediation, lowering investigation time from hours to minutes. This resolution not solely reduces the amount of assist requests and accelerates the time to decision, but additionally unlocks larger operational effectivity for finish customers.

By reimagining how customers interact with the system—from AI-powered workflows to smarter automation—Druva noticed a transparent alternative to ship a extra seamless buyer expertise that strengthens buyer satisfaction, loyalty, and long-term success.

The important thing alternatives for Druva in implementing a generative AI-powered multi-agent copilot embody:

• Simplified consumer expertise: By offering a pure language interface, the copilot can simplify advanced knowledge safety duties and assist customers entry the data they want rapidly.
• Clever Troubleshooting: The copilot can leverage AI capabilities to research knowledge from varied sources, determine the foundation causes of backup failures, and supply personalised suggestions for decision.
• Streamlined Coverage Administration: The multi-agent copilot can information customers by way of the method of making, modifying, and implementing knowledge safety insurance policies, lowering the potential for human errors and bettering compliance.
• Proactive Assist: By constantly monitoring knowledge safety environments, the copilot can proactively determine potential points and supply steerage to assist forestall failures or optimize efficiency.
• Scalable and Environment friendly Operations: The AI-powered resolution can deal with a big quantity of buyer inquiries and duties concurrently, lowering the burden on Druva’s assist crew in order that they’ll concentrate on extra advanced and strategic initiatives.

Resolution overview

The proposed resolution for Druva’scopilot leverages a classy structure that mixes the ability of Amazon Bedrock (together with Amazon Bedrock Data Bases), LLMs, and a dynamic API choice course of to ship an clever and environment friendly consumer expertise. Within the following diagram, we exhibit the end-to-end structure and varied sub-components.

On the core of the system is the supervisor agent, which serves because the central coordination element of the multi-agent system. This agent is answerable for overseeing your entire dialog circulation, delegating duties to specialised sub-agents, and sustaining seamless communication between the varied parts.

The consumer interacts with the supervisor agent by way of a consumer interface, submitting pure language queries associated to knowledge safety, backup administration, and troubleshooting. The supervisor agent analyzes the consumer’s enter and routes the request to the suitable sub-agents primarily based on the character of the question.

The knowledge agent is answerable for retrieving related info from Druva’s programs by interacting with the GET APIs. This agent fetches knowledge resembling scheduled backup jobs, backup standing, and different pertinent particulars to supply the consumer with correct and up-to-date info.

The assistance agent assists customers by offering steerage on greatest practices, step-by-step directions, and troubleshooting suggestions. This agent attracts upon an in depth data base, which incorporates detailed API documentation, consumer manuals, and ceaselessly requested questions, to ship context-specific help to customers.

When a consumer must carry out vital actions, resembling initiating a backup job or modifying knowledge safety insurance policies, the motion agent comes into play. This agent interacts with the POST API endpoints to execute the mandatory operations, ensuring that the consumer’s necessities are met promptly and precisely.

To guarantee that the multi-agent copilot operates with essentially the most appropriate APIs and parameters, the answer incorporates a dynamic API choice course of. Within the following diagram, we spotlight the varied AWS companies used to implement dynamic API choice, with which each the info agent and the motion agent are outfitted. Bedrock Data Bases comprises complete details about out there APIs, their functionalities, and optimum utilization patterns. As soon as an enter question is acquired, we use semantic search to retrieve the highest Ok related APIs. This semantic search functionality permits the system to adapt to the precise context of every consumer request, enhancing the Copilot’s accuracy, effectivity, and scalability. As soon as the suitable APIs are recognized, the agent prompts the LLM to parse the highest Ok related APIs and finalize the API choice together with the required parameters. This step makes certain that the copilot is absolutely outfitted to run the consumer’s request successfully.

Lastly, the chosen API is invoked, and the multi-agent copilot carries out the specified motion or retrieves the requested info. The consumer receives a transparent and concise response, together with related suggestions or steerage, by way of the consumer interface.

All through the interplay, customers can present further info or specific approvals by utilizing the consumer suggestions node earlier than the copilot performs vital actions. With this human-in-the-loop strategy, the system operates with the mandatory safeguards and maintains consumer management over delicate operations.

Analysis

The analysis course of for Druva’s generative AI-powered multi-agent copilot focuses on assessing the efficiency and effectiveness of every vital element of the system. By totally testing particular person parts resembling dynamic API choice, remoted checks on particular person brokers, and end-to-end performance, the copilot delivers correct, dependable, and environment friendly outcomes to its customers.

Analysis methodology:

• Unit testing: Remoted checks are carried out for every element (particular person brokers, knowledge extraction, API choice) to confirm their performance, efficiency, and error dealing with capabilities.
• Integration Testing: Exams are carried out to validate the seamless integration and communication between the varied parts of the multi-agent copilot, sustaining knowledge circulation and management circulation integrity.
• System Testing: Finish-to-end checks are executed on the entire system, simulating real-world consumer situations and workflows to evaluate the general performance, efficiency, and consumer expertise.

Analysis outcomes

Selecting the best mannequin for the fitting activity is vital to the system’s efficiency. The dynamic device choice represents probably the most vital components of the system—invoking the right API is crucial for end-to-end resolution success. A single incorrect API name can result in fetching mistaken knowledge, which cascades into inaccurate outcomes all through the multi-agent system. To optimize the dynamic device choice element, varied Nova and Anthropic fashions have been examined and benchmarked in opposition to the bottom fact created utilizing Sonnet 3.7.

The findings confirmed that even smaller fashions like Nova Lite and Haiku 3 have been in a position to choose the right API each time. Nevertheless, these smaller fashions struggled with parameter parsing resembling calling the API with the right parameters relative to the enter query. When parameter parsing accuracy was taken under consideration, the general API choice accuracy dropped to 81% for Nova Micro, 88% for Nova Lite, and 93% for Nova Professional. The efficiency of Haiku 3, Haiku 3.5, and Sonnet 3.5 was comparable, starting from 91% to 92%. Nova Professional offered an optimum tradeoff between accuracy and latency with a mean response time of simply over one second. In distinction, Sonnet 3.5 had a latency of eight seconds, though this may very well be attributed to Sonnet 3.5’s extra verbose output, producing a mean of 291 tokens in comparison with Nova Professional’s 86 tokens. The prompts might doubtlessly be optimized to make Sonnet 3.5’s output extra concise, thus lowering the latency.

For end-to-end testing of actual world situations, it’s important to interact human material skilled evaluators aware of the system to evaluate efficiency primarily based on completeness, accuracy, and relevance of the options. Throughout 11 difficult questions in the course of the preliminary improvement part, the system achieved scores averaging 3.3 out of 5 throughout these dimensions. This represented stable efficiency contemplating the analysis was carried out within the early levels of improvement, offering a robust basis for future enhancements.

By specializing in evaluating every vital element and conducting rigorous end-to-end testing, Druva has made certain that the generative AI-powered multi-agent copilot meets the best requirements of accuracy, reliability, and effectivity. The insights gained from this analysis course of have guided the continual enchancment and optimization of the copilot.

“Druva is on the forefront of leveraging superior AI applied sciences to revolutionize the way in which organizations defend and handle their vital knowledge. Our Generative AI-powered Multi-agent Copilot is a testomony to our dedication to delivering revolutionary options that simplify advanced processes and improve buyer experiences. By collaborating with the AWS Generative AI Innovation Middle, we’re embarking on a transformative journey to create an interactive, personalised, and environment friendly end-to-end expertise for our clients. We’re excited to harness the ability of Amazon Bedrock and our proprietary knowledge to proceed reimagining the way forward for knowledge safety and cyber resilience.”- David Gildea, VP of Generative AI at Druva

Conclusion

Druva’s generative AI-powered multi-agent copilot showcases the immense potential of mixing structured and unstructured knowledge sources utilizing AI to create next-generation digital copilots. This revolutionary strategy units Druva other than conventional knowledge safety distributors by reworking hours-long guide investigations into on the spot, AI-powered conversational insights, with 90% of routine knowledge safety duties executable by way of pure language interactions, essentially redefining buyer expectations within the knowledge safety house. For organizations within the knowledge safety and safety house, this expertise permits extra environment friendly operations, enhanced buyer engagement, and data-driven decision-making. The insights and intelligence offered by the copilot empower Druva’s stakeholders, together with clients, assist groups, companions, and executives, to make knowledgeable selections sooner, lowering common time-to-resolution for knowledge safety points by as much as 70% and accelerating backup troubleshooting from hours to minutes. Though this venture focuses on the info safety trade, the underlying ideas and methodology could be utilized throughout varied domains. With cautious design, testing, and steady enchancment, organizations in any trade can profit from AI-powered copilots that contextualize their knowledge, paperwork, and content material to ship clever and personalised experiences.

This implementation leverages Amazon Bedrock AgentCore Runtime and Amazon Bedrock AgentCore Gateway to supply strong agent orchestration and administration capabilities. This strategy has the potential to supply clever automation and knowledge search capabilities by way of customizable brokers, reworking consumer interactions with functions to be extra pure, environment friendly, and efficient. For these excited about implementing comparable functionalities, discover Amazon Bedrock Agents, Amazon Bedrock Knowledge Bases and Amazon Bedrock AgentCore as a completely managed AWS resolution.

In regards to the authors

David Gildea With over 25 years of expertise in cloud automation and rising applied sciences, David has led transformative initiatives in knowledge administration and cloud infrastructure. Because the founder and former CEO of CloudRanger, he pioneered revolutionary options to optimize cloud operations, later resulting in its acquisition by Druva. Presently, David leads the Labs crew within the Workplace of the CTO, spearheading R&D into Generative AI initiatives throughout the group, together with initiatives like Dru Copilot, Dru Examine, and Amazon Q. His experience spans technical analysis, industrial planning, and product improvement, making him a distinguished determine within the subject of cloud expertise and generative AI.

Tom Nijs is an skilled backend and AI engineer at Druva, pushed by a ardour for each studying and sharing data. Because the Lead Architect for Druva’s Labs crew, he channels this ardour into growing cutting-edge options, main initiatives resembling Dru Copilot, Dru Examine, and Dru AI Labs. With a core concentrate on optimizing programs and harnessing the ability of AI, Tom is devoted to serving to groups and builders flip groundbreaking concepts into actuality.

Gauhar Bains is a Deep Studying Architect on the AWS Generative AI Innovation Middle, the place he designs and delivers revolutionary GenAI options for enterprise clients. With a ardour for leveraging cutting-edge AI applied sciences, Gauhar makes a speciality of growing agentic AI functions, and implementing accountable AI practices throughout numerous industries.

Ayushi Gupta is a Senior Technical Account Supervisor at AWS who companions with organizations to architect optimum cloud options. She makes a speciality of guaranteeing business-critical functions function reliably whereas balancing efficiency, safety, and value effectivity. With a ardour for GenAI innovation, Ayushi helps clients leverage cloud applied sciences that ship measurable enterprise worth whereas sustaining strong knowledge safety and compliance requirements.

Marius Moisescu is a Machine Studying Engineer on the AWS Generative AI Innovation Middle. He works with clients to develop agentic functions. His pursuits are deep analysis brokers and analysis of multi agent architectures.

Ahsan Ali is an Senior Utilized Scientist on the Amazon Generative AI Innovation Middle, the place he works with clients from completely different trade verticals to resolve their pressing and costly issues utilizing Generative AI.

Sandy Farr is an Utilized Science Supervisor on the AWS Generative AI Innovation Middle, the place he leads a crew of scientists, deep studying architects and software program engineers to ship revolutionary GenAI options for AWS clients. Sandy holds a PhD in Physics and has over a decade of expertise growing AI/ML, NLP and GenAI options for giant organizations.

Govindarajan Varadan is a Supervisor of the Options Structure crew at Amazon Internet Companies (AWS) primarily based out of Silicon Valley in California. He works with AWS clients to assist them obtain their enterprise targets by way of revolutionary functions of AI at scale.

Saeideh Shahrokh Esfahani is an Utilized Scientist on the Amazon Generative AI Innovation Middle, the place she focuses on reworking cutting-edge AI applied sciences into sensible options that tackle real-world challenges.