In The News
-Sathish Raman
AI agents are beginning to take on work that has traditionally required people: making calls, following up with carriers, checking eligibility, gathering information and handling routine operational decisions. But getting an AI system to complete a task is only part of the challenge. Making it fast, reliable and responsive enough to operate inside a real business environment requires a different layer of engineering.
Dhaarna Maheshwari is the engineering force behind HappyRobot’s robust AI agents. She builds the critical infrastructure for real-time context, model integration, and data management, ensuring AI systems perform reliably at scale. Discover how her innovative solutions make HappyRobot’s AI agents operate seamlessly in complex, dynamic business environments.
That is where software engineer Dhaarna Maheshwari has focused her work at HappyRobot, an AI company building voice and text agents for supply-chain operations. Her work spans the infrastructure underneath those agents – from real-time context and model integrations to performance, reliability and data management.
A major part of that work has been making the platform’s agent infrastructure more unified and scalable. As HappyRobot expanded its AI capabilities, Maheshwari worked across the underlying systems that power its voice and text agents, helping consolidate shared functionality and modernize the platform’s LLM infrastructure. This included moving the company’s integration toward OpenAI’s newer APIs and building the supporting infrastructure needed to operate those models reliably at scale.
One of the more important pieces of that infrastructure is Signals, a system designed to give agents access to new information while an interaction is already underway. In a supply-chain workflow, information can change in seconds: a carrier may respond to an email, a shipment status may be updated or another part of the system may produce new information while an agent is still on a call.
Signals allows that information to reach the agent during the interaction rather than leaving it dependent on the context available when the conversation began. The result is an agent that can respond to events as they happen instead of operating from a static snapshot of information.
That same focus on real-time behavior extended to HappyRobot’s migration to OpenAI’s Responses API. The migration exposed performance issues that were not immediately obvious. Some workloads became slower and more expensive despite the move to newer infrastructure.
Maheshwari investigated the behavior at the systems level and traced the problem to how previous AI work was being reused. Addressing the underlying issue improved performance and reduced costs for workloads running on the newer models, with some models ultimately outperforming the company’s previous setup. She also identified a separate performance issue affecting older models, documented the findings for OpenAI and helped strengthen monitoring so similar problems could be detected earlier.
Model performance, however, is only one part of making an AI agent reliable in production. Maheshwari also worked on output safeguards designed to prevent unintended model responses from reaching customers. AI-generated output can occasionally contain duplicated messages, malformed language, internal tool syntax or other artifacts that make an automated interaction feel unreliable. The platform therefore applies checks and cleanup before raw model output is passed through to the customer-facing experience.
Behind those agent interactions is another infrastructure challenge: data growth. One of HappyRobot’s databases had grown beyond six terabytes while continuing to accumulate tens of gigabytes of data each day. Keeping all of that information in the primary database would increase storage costs and make the system increasingly difficult to operate efficiently.
Maheshwari designed and led the development of an archival system that moves older data into lower-cost storage while preserving access when that historical information is needed. The approach allows the primary database to remain manageable without sacrificing the ability to retrieve older records.
Together, these projects reflect a broader engineering challenge in production AI: the model itself is only one component of the system. Real-world agents need infrastructure that can deliver fresh information during an interaction, integrate changing model capabilities, control latency and cost, handle unexpected output and continue operating as the volume of data grows.
That focus builds on Maheshwari’s earlier work across AI and large-scale software systems. Before joining HappyRobot, she worked at Goldman Sachs, C3.ai and Millennium Management, where she contributed to an internal LLM and retrieval-augmented generation platform used by developers and investment professionals. She holds a master’s degree in computer science from Columbia University, with a focus on machine learning.
At HappyRobot, the work sits largely behind the interface that customers see. The visible product is an AI agent having a conversation or completing a workflow. Underneath it is a much larger engineering system responsible for deciding what information the agent sees, how quickly it can respond, what happens when a model behaves unexpectedly and how the platform continues to operate as usage grows.
For AI agents moving from demonstrations into real operational environments, that infrastructure is becoming a critical part of the product itself.
