Divakar Kumar

Divakar Kumar builds AI applications and teaches architectures for agents that interact with desktops, retrieve organizational knowledge, and act on event streams. His speaker profile identifies him as a technical architect at FlyersSoft, a Microsoft MVP in AI, and a Microsoft Certified Trainer.[2][4][5]

Kumar’s financial-agent architecture starts with event sourcing: transactions, logins, and transfers are preserved as immutable events that an agent can correlate with historical activity. His session “Agentic Event-Driven Systems: When Events Talk Back” describes agents generating new domain events, such as suspicious-activity flags or transactions held for review. It also addresses preserving system invariants and preventing runaway feedback loops among fraud, compliance, and customer-service agents. Latent Space’s July 29, 2026 finance-track coverage characterized his argument as using existing event histories to underpin auditable production decision loops.[2][1]

In November 2025, Kumar documented Cuata, a computer-using agent he built for briefly missed meeting content. The starting problem was practical: someone steps away for a doorbell, an urgent call, or coffee and wants to return to a summary with screenshots of key slides. He described extending the same system to browsing, reading articles, and writing summaries into Microsoft Word.[4]

Kumar organized Cuata around a think → select strategy → execute loop, with Semantic Kernel coordinating ten plugins in his November 2025 design. Azure OCR locates screen text and supplies coordinates for mouse actions; screenshot analysis checks whether the expected result appeared. In his Teams example, the agent checks that the calendar loaded, locates the meeting, clicks it, and validates each subsequent step. A failed action can lead to a retry, a different search term, or a request for help.[4]

He chose desktop-level interaction to support workflows spanning applications: copying from Chrome into Word, sending content through Outlook, and interacting with desktop dialogs. Mouse and keyboard simulation provide the actions, while OCR and Azure OpenAI vision provide screen interpretation. The screenshot checks are part of his described feedback mechanism; the article supplies no measured reliability result.[4]

On February 21, 2026, Kumar explained his decision to reduce GitHub Copilot Agent mode use: “I didn’t get the satisfaction that I had built something myself, even though productivity increased.” He described switching to Ask mode for questions and debugging, retaining inline suggestions for boilerplate, and occasionally using Agent mode for well-defined tasks whose solutions he had already considered. He framed this as personal experience rather than general advice, emphasizing learning, thinking through edge cases, and ownership of the work.[3]

Before Cuata and his 2026 agent discussions, Kumar’s February 2024 CosmicTalent account documented AI retrieval work for FlyersSoft. He described developing an application to help HR and managers find employees suited to task requirements. That earlier work provides a concrete example of his decisions about preparing organizational data for language-model responses.[5]

The initial version embedded whole resumes and encountered retrieval problems and token limits. Kumar described the team’s refinement: a custom Azure Document Intelligence extraction model produced structured resume fields, which were split into chunks. Each chunk repeated the employee’s ID and name to retain identity context. Only project details were summarized before embedding; other fields bypassed summarization. Embeddings were stored in Azure Cosmos DB Mongo vCore, and retrieved records were passed to GPT-3.5 Turbo to generate answers. Kumar reported better retrieval and token consumption without quantifying the improvements.[5]