DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES
Ph.D. Thesis Colloquium
Speaker : Mr. Varad Kulkarni Vinod
S.R. Number : 06-18-01-10-12-21-1-19680
Title : “Serverless Orchestration of Classical and Agentic Workflows on Multi-Cloud FaaS Platforms”
Research Supervisor : Prof. Yogesh Simmhan
Date & Time : July 29, 2026 (Wednesday), 16:00 PM
Venue : #102, CDS Seminar Hall
ABSTRACT
Functions-as-a-Service (FaaS) has become a natural building block for elastic, event-driven cloud applications: developers write short-lived functions, and the platform handles scaling and billing. As applications grow beyond single functions, providers expose workflow orchestrators, such as AWS Step Functions and Azure Durable Functions, that compose these functions into multi-step pipelines. At the same time, a new class of applications is emerging: agentic workflows, in which large language model (LLM) agents iteratively reason, call tools, and adapt their control flow based on intermediate progress. Orchestrating both classical and agentic workflows efficiently across multiple clouds remains difficult. Provider platforms are proprietary and asymmetric; placement and fusion decisions lack principled multi-cloud algorithms; and the stateful, iterative nature of agentic patterns sits uneasily with the largely stateless FaaS execution model.
This talk presents a systems stack for serverless orchestration of classical and agentic workflows on multi-cloud FaaS platforms, and closes the loop with a live application of that stack in graduate education.
We begin by asking what modern FaaS workflow platforms actually cost and where time goes. Through XFBench, a cross-cloud benchmark suite spanning AWS and Azure, we characterize orchestration and inter-function communication as first-class performance factors, and show that cold starts and topology interact in ways that single-function benchmarks miss. Building on these asymmetries, we design XFaaS: a portable compilation and optimization layer that deploys provider-independent workflows onto native AWS and Azure orchestrators, and jointly decides where functions should run and which adjacent steps should be fused. The goal is not a single “best cloud,” but principled hybrid-cloud planning under latency and cost objectives.
The second half of the talk turns to agentic workloads. We introduce FAME, a middleware that maps common agent patterns onto serverless orchestration with durable memory and Model Context Protocol (MCP) tooling, so iterative LLM agents can run without reinventing state and tool plumbing for each cloud. We then extend portability to multi-cloud agentic settings with XFAGENT, which supports conditional loops and declarative agent configuration while generating native orchestration code across providers. Together, these layers treat agentic applications as first-class citizens of the FaaS workflow ecosystem rather than ad hoc scripts around chat APIs.
Finally, we validate the stack outside micro-benchmarks by deploying it in a live graduate Cloud Computing course at IISc. An in-class AI Tutor Agent is coupled with Moodle-triggered FaaS evaluation workflows that generate student feedback and instructor analytics under the bursty submission patterns of real lectures. Alongside a systems evaluation of these classroom workflows, we report pedagogical insights from linking tutor transcripts to proctored assessment: how students engage with the agent, where transcript coverage aligns with quiz performance, and where discrimination among related concepts still breaks down despite exposure. Beyond showing that the systems stack is operationally viable for education, the deployment raises a deeper question for AI-native applications: scalable orchestration is necessary, but application-level alignment between dialogue, analytics, and assessment still determines whether the system delivers meaningful outcomes.
Overall, the talk argues that classical and agentic multi-cloud FaaS orchestration can be made portable, optimizable, and practically deployable, and that closing the loop with a real classroom application is as important as building the infrastructure itself.
ALL ARE WELCOME
