IEEE SCV WIE AI Summit 2026

Neeraja of Workday

As generative AI models become embedded in daily software engineering, the biggest lever teams overlook isn't which model to use — it's how efficiently they feed it. Every unnecessary token in a prompt or context window carries a cost: in latency, in inference spend, and in output reliability, as irrelevant context measurably increases hallucination rates.

Biography:

Neeraja Amilineni is a Senior Software Development Engineer at Workday, based in Pleasanton, CA, with 15 years of hands-on experience building and scaling enterprise software systems. She designs and develops core architecture, leads complex technical initiatives, performs code reviews, and mentors fellow engineers to drive technical excellence across applications and services. Today, she focuses on helping engineering teams work effectively with generative AI.

Outside of engineering, Neeraja is a proud mom of two teenage daughters. She loves reading autobiographies, exploring different meditation practices, and has recently discovered a passion for photography. She also picks up a new skill every year or two, from swimming to dance, bringing the same curiosity and growth mindset to her personal life that she brings to her engineering work.

Mayuri Kulkarni

Enterprise networks have spent two decades getting more programmable — SDN, intent-based networking, telemetry-driven automation — yet day-to-day operations still lean heavily on human engineers to interpret dashboards, correlate alerts, and execute runbooks. Agentic NetOps changes that equation: AI agents that reason over network state, make context-aware decisions, and autonomously execute multi-step operational workflows, rather than simply flagging anomalies for a human to resolve.

Drawing on firsthand experience building AI-driven networking products, this session breaks the shift into three parts. First, what actually makes an operation "agentic" — the distinction between scripted automation, AI-assisted troubleshooting, and true autonomous decision-making with guardrails. Second, where agentic NetOps delivers the most value today: WAN and connectivity fabric management, branch and campus operations, security policy enforcement, and hybrid cloud networking — domains with high operational toil and well-structured telemetry. Third, a practical framework for building enterprise-grade agentic solutions: the data and observability foundation required, how to scope agent autonomy safely, how to build eval frameworks, human-in-the-loop checkpoints, and the trust and verification layers needed before agents can act on production infrastructure.

Attendees will leave with a clear mental model of agentic NetOps, a practical lens for spotting where it applies in their own environment, and a blueprint for piloting it responsibly — grounded in real lessons from building these systems.

Biography:

Mayuri Kulkarni is a product executive with over 20 years of experience across Product Management and Engineering in Enterprise Network Security and SASE space. She has led products through their full lifecycle — from new-product incubation to market leadership — across the full Enterprise Networking stack, including SD-WAN, Data Center Networking, LAN, NAC, and Security. She spent part of her career at Cisco Systems, where she contributed to the company's core data center networking portfolio, and more recently at Palo Alto Networks, where she led Prisma SD-WAN Software and AI and Agentic AI innovations for SASE Networking use cases, building market differentiation of the product. Her technical background spans data center switching and routing, application acceleration, and network security software. She has been a trusted advisor to enterprise customers navigating their digital transformation and AI adoption journey, and has mentored and grown product teams.

Niruta Talwekar

Teams are increasingly wiring LLMs into every stage of extraction pipelines: extract with a model, classify with a model, check with a model, then route the doubtful cases to a human. The appealing version of that last step is a judge panel — run several models over each field and treat disagreement as the signal for human review. No ground truth needed at inference time. But the more transferable finding is how hard it was to learn that. This measurement is unusually good at producing wrong numbers that look right — nine distinct failure modes, six of which made the results look better rather than worse. This talk walks the three questions in order — can you trust your measurement, can you trust your panel, can you trust your label space — and ends with a checklist of five audits that no standard evaluation practice covers.

Biography:

Niruta Talwekar is focused on keeping human judgment in the loop as AI takes on an increasingly larger role in development. As a Staff Data Engineer on Meta's Experimentation Platform team, she builds the infrastructure that measures whether products and AI features actually deliver results—along with the robust data quality checks that catch LLM pipeline failures when they fall short. While AI has made building easier than ever, Niruta ensures we have the right tools to measure what truly succeeds.

Arpita

LLM-powered products increasingly show users the reasoning behind an output — a recommendation, a ranking, a decision. That display carries an implicit claim: this is why the system chose what it chose. This talk asks whether that claim holds.

The method is a counterfactual audit. We take the model's own generated reasoning and edit it with four operators: flip the stated preference, swap in a different user's reasoning, erase it entirely, and paraphrase it as a control. We then measure how often the top-1 output changes. The faithfulness gap — the average change rate under the three meaning-changing edits, minus the change rate under paraphrasing — separates genuine reason-responsiveness from ordinary output instability.

Applied to GPT-5 Chat, GPT-5 Reasoning, and DeepSeek-R1 on a public Amazon dataset, the results diverge sharply. GPT-5 Chat with chain-of-thought prompting is strongly reason-responsive: flipping its stated preference changes the top recommendation for 94% of users, while paraphrasing changes it for 5%. The two native reasoning models show almost no such signal in their displayed explanations, with gaps of 0.02 and −0.11. When the same audit is applied to DeepSeek-R1's hidden reasoning trace instead, the gap rises to 0.30 — the output tracks the internal trace more closely than the explanation shown to the user.

The audit is training-free, model-agnostic, and inexpensive to run. Attendees will leave able to apply it to their own systems before shipping user-facing explanations. 

Biography:

Arpita Vasant Shah is a Senior Applied Scientist at Microsoft, working
on LLMs for ranking and personalization in production. She has
thirteen years in industry,,previously at Amazon Lab126 and Amazon
Music, and advises on a strategic AI course at the University of San
Francisco. Her work on auditing LLM reasoning faithfulness was
accepted at ACM RecSys 2026. She is an IEEE Senior Member and reviews
for IEEE TNNLS, IJCNN, AAAI, and NeurIPS.

Negin Heidari

Robots that manipulate objects, carry out tasks, and interact with people offer an exciting glimpse into physical AI. Behind these demonstrations lies an engineering question: how did we move from models that predict words to systems that perceive, decide, and act?

Drawing on publicly available research and representative robotics examples, this session traces the evolution from Large Language Models (LLMs) to Vision-Language Models (VLMs), Vision-Language-Action models (VLAs), World Action Models, and embodied physical AI. It explores how language, perception, and reasoning connect to control loops for physical action, and how real-time interaction, memory, and world prediction support continuous observation–action feedback.

The session also examines the ethical responsibility of developing systems whose decisions have physical consequences. Using the trolley problem as a discussion lens, it considers how bias, assumptions, and competing values influence automated behavior, and why exciting demonstrations require careful evaluation and accountable decisions.

Attendees will gain an understanding of the evolution toward physical intelligence and our responsibilities as developers, engineers, and technology leaders shaping the world these systems inhabit.

Biography:

Negin Heidari is an Embodied AI and Control Engineer at SF Motors, with interests in vision-language-action models, multimodal AI, and embodied intelligence for robotics.

Her background spans undergraduate control systems and graduate studies in communications and AI. She earned her M.S. in Electrical and Computer Engineering at Santa Clara University, where she worked as a Graduate Research Assistant in the Information Theory and Machine Learning Laboratory on retrieval algorithm optimization, and in the Machine Learning and Safety Analytics Laboratory on robotic perception and audio-based human–robot interaction.

An IEEE officer in Santa Clara Valley with leadership roles across multiple chapters, Negin finds physical AI especially exciting for bringing together control, communication, perception, and reasoning. Her interests include both its possibilities and the ethical responsibility of developing systems that act in the world.