Chibi Vikram

Chibi Vikram helped build the evaluation pipelines that test UiPath’s AI agents and the serverless systems that run them for enterprise customers. On the evaluations team, Chibi and his colleagues worked toward a common way to test agents, whether a builder assembled one in a visual editor or wrote it in code.

One problem the team addressed was how to spot an instruction that no test had exercised. Suppose a document-processing agent has instructions to ask for a missing field. If every test document is complete, the agent can pass every test without ever asking for the field. The team’s prompt coverage work makes that gap visible. Adding an incomplete document exercises the missing case. The builder still needs to decide what counts as an appropriate response. For agents written in code, Chibi proposed a proof of concept for custom Python evaluators so builders could express those criteria in their own code.

Chibi also helped integrate the Jev model into Flow, UiPath’s visual workflow editor, where builders can use it directly as a step in an agent’s workflow. The team’s J Stack includes JExtract, which turns a document into structured fields for a later step, and JRank, which takes results a retrieval system has already found and decides which should come first.

Shared controls and an in-app assistant

Chibi helped design a shared Web Components system for UiPath products. He wanted teams to be able to reuse the way a control behaves as well as the way it looks. By keeping interactions inside the component, teams could use the same control across products without implementing those interactions again.

Chibi was one of five named co-inventors on UiPath’s digital assistant patent application, published in February 2024. In the proposed design, a person’s action inside an application prompts the assistant to offer relevant automations. When the person chooses one, a side panel updates with suggested follow-up actions. The person can discover an automation while working on the task where it applies.

Calling tools from a web chat

Chibi’s open-source Next.js MCP Client lets a developer import a Claude Desktop configuration, discover the tools its Model Context Protocol servers offer, and call them from a web chat. Where those servers run matters when deploying the client. A serverless Next.js deployment can connect to a remote server. If the imported configuration calls for launching a local server process, the developer needs a persistent container to run it.