Divakar Kumar

Divakar Kumar is a technical architect at Flyers Soft, a Microsoft MVP in AI, and a Microsoft Certified Trainer who builds AI agents for established enterprise systems. His work addresses a central challenge of deploying autonomous software in consequential environments: giving agents enough context to make useful decisions without sacrificing operational boundaries, reliability, or human judgment.

Kumar joined Flyers Soft in May 2023. His earlier public projects include Azure Functions middleware and an Alexa-powered presentation controller. By January 2025, he was developing durable multi-agent architectures combining Semantic Kernel, Azure Durable Functions, and distributed-systems orchestration.

• Context-aware fraud detection. Kumar’s event-sourced fraud architecture reserves AI agents for ambiguous transactions while existing rules and machine-learning models handle routine decisions. Azure Cosmos DB, change feeds, and materialized views assemble transaction histories, device fingerprints, customer-verification records, and payment signals without dissolving domain boundaries. Specialized risk and behavior agents assess each case before a final verdict returns to the event-driven workflow; bounded reasoning and latency constraints keep the system operationally practical.

• Cuata, a computer-using assistant. Kumar built Cuata around Semantic Kernel and plugins for interpreting screens, operating desktop applications, and validating actions. Its documented workflows include attending Microsoft Teams meetings, capturing and summarizing missed material, researching through a browser, and writing findings into Microsoft Word.

• Graph-grounded event intelligence. His EventKernel and GraphRAG architecture combines vector retrieval, Neo4j graph traversal, and deterministic queries to recommend conference sessions using relationships among speakers, attendees, previous interactions, and interests.

• Developer agency over indiscriminate automation. After stepping back from GitHub Copilot’s Agent mode, Kumar argued that autonomous coding can undermine learning, ownership, and technical judgment. He favors inline assistance and question-answering for everyday development, reserving autonomous execution for well-defined tasks while accounting for unnecessary computational costs.