Aman Mundra
Founder & CEO, Welzin | Co-founder & CEO, CogNerd
Chandigarh, India
Actions
Aman Mundra (Amandeep Singh) is Founder & CEO of Welzin, a full-stack AI and data firm, and Co-founder & CEO of CogNerd, a platform that tracks how AI engines like ChatGPT and Perplexity see brands.
He spent 12+ years building ML systems at PayPal, Koch Industries, TransUnion, HPE and HealthifyMe - including risk, fraud and payment optimization models adding ~$5M/yr to PayPal's authorization rates - with roots in Spark data engineering and recent work in GenAI/RAG, LLM evaluation and fine-tuning.
Where most AEO voices come from SEO, his comes from production ML: he builds the instruments that measure how AI engines describe brands. He holds a B.Tech in Computer Science (2010–14), has spoken at KubeCon + CloudNativeCon India 2025, Open Source Summit Japan, KCD Sri Lanka and KubeVirt Summit, KCD-Delhi, and writes at amanmundra.com.
Area of Expertise
Topics
Inference at the Edge: Running AI on Microcontrollers with Zephyr RTOS
Microcontrollers have no GPU, no cloud connection, and sometimes less than 256KB of RAM. TinyML is transforming edge computing by enabling smart inference directly on microcontrollers.
Zephyr RTOS, with its lightweight footprint, modular build system, and growing hardware support, has become one of the most capable platforms for building production-grade embedded AI systems.
This session is a practical, hardware-grounded walkthrough of what it actually takes to deploy and optimize TinyML workloads on Zephyr.
We will cover how to evaluate and select the right inference runtime for your hardware constraints, comparing TensorFlow Lite Micro, microTVM, emlearn, and LiteRT across memory footprint, operator support, and ease of integration.
We will walk through Zephyr's Linkable Loadable Extensions (LLEXT), which allow models to be hot-swapped at runtime without reflashing the device, a critical capability for OTA model updates in the field.
On the optimization side, we will go hands-on with quantization and operator fusion, and benchmark the results on physical hardware against Renode simulation, so you understand exactly where simulation diverges from reality.
Breaking RAG Apart: A Microservices Blueprint for Retrieval at Scale on Kubernetes
Most RAG systems are just jupyter notebook POCs: embed, retrieve, generate, all in a few cells that can hardly support 10 concurrent queries. But with the real traffic, Vector search fails on high concurrency, the GPU generation step becomes a bottleneck because you never tuned it, and one slow re-rank call takes the whole request down with it.
This talk covers how to split a RAG pipeline into separate services on Kubernetes: embedding, vector retrieval, re-ranking, orchestration, LLM generation and evaluation. Each of these has a different scaling profile and a different way of failing, so each needs to be deployed and scaled on its own terms. I'll walk through running vector databases as sharded, stateful services, autoscaling GPU inference with KEDA instead of relying on CPU-based HPA, where caching actually pays off, and how to keep one failing component from taking down the whole system.
Attendees will have a working reference architecture for a scalable RAG pipeline that serve thousands of concurrent queries, along with a clear picture of which Kubernetes building blocks (StatefulSets, operators, KEDA, GPU scheduling) map to which part of the pipeline. Along with that, a bonus architecture pipeline for performing evaluation at scale.
How to do deep contextual tracing of Agentic decisions with OpenSearch
AI agents rely on LLM reasoning or vector similarity to decide and choose tools. But these approaches lack deterministic control, explainability and low-latency guarantees. In real production systems, agents need a fast and reliable way to decide which tool to invoke based on structured rules and historical context.
With this, Agentic observability has emerged as a critical discipline in 2026, shifting from simple logging to in-depth tracing of non-deterministic, multi-step workflows. Building a reliable GenAI app monitoring strategy for tracing the reasoning chain, now becomes more important from an observability pov.
In this talk, we will demonstrate how we can understand AI Agent decisioning for tool calling and further reasoning with OpenSearch. We will showcase how OpenSearch indexes agent logs, tool call metadata, execution constraints and past outcomes.
We will showcase a live hands-on demo of OpenSearch ML, indexing the AI Agent logs-metrics and further use them to comprehend the logs to understand the functioning and reasoning of AI agents.
Confused about how AI Agents are making decisions ? Then this talk is for you !
Deep Contextual Tracing of Agentic Decisions with Apache Lucene
AI agents often rely on LLM reasoning or vector similarity to choose tools, but these approaches lack deterministic control, explainability and low-latency guarantees. In real production systems, agents need a fast and reliable way to decide which tool to invoke based on structured rules and historical context.
In this talk, we will demonstrate how Apache Lucene can be used to understand AI Agents decisioning for tool calling and output generation. We will showcase how Lucene indexes agent logs, tool metadata, execution constraints and past outcomes. We will also show how Lucene can be used to filter, analyze and score agent tool calls at runtime.
We will show a live hands-on demo of how Apache Lucene can be used to store logs-metrics and how to comprehend these logs to understand the functioning of AI agents.
Attendees will also get GitHub templates and code samples for the complete demo implementation.
LLMs as a Judge: using LLMs and Evaluation frameworks for model improvement
In this 1.5 hours hands on workshop, you will learn how to design and implement a evaluative LLMOps pipeline. You will learn how to Implement feedback loop from live monitoring, back into the evaluation pipeline.
We will do a hands on workshop on a real world example of a RAG pipeline and integrate open source evaluation framework like RAGas, Evidently, Langsmith and Opik.
We will also demonstrate how to rigorously evaluate the model outputs, monitor their behavior and implement human in the loop assessment for continuous model improvement.
We will also cover:
1. Why LLM as a Judge works and when (and when not) to use it.
2. How to write evaluation prompts for binary scoring, chain of thoughts and structured output.
3. How to manage bias, verify against human ground truth and pitfalls in scaling and implementing evaluation methods in LLMOps pipeline.
KMCP and the Future of Agent-Native Kubernetes Control Planes
As AI agents become part of production platforms, their interaction with Kubernetes is increasingly ad hoc. Each agent embeds its own assumptions, permissions, and execution logic, making behaviour difficult to reason about, audit, or operate safely at scale.
KMCP introduces a missing control plane interface, a standardized contract for how agents exchange intent, context, and actions with Kubernetes systems.
This session explores how KMCP can be used as an infrastructure primitive rather than an AI feature. We’ll walk through practical platform scenarios where KMCP sits between agent logic and Kubernetes APIs, defining clear boundaries for execution, validation, and observability.
In this session, you’ll see how KMCP enables platform teams to :
Standardize agent-to-cluster interactions across tools, controllers, and workflows
Separate agent intent from execution logic, reducing tight coupling with Kubernetes internals
Enforce permission and policy boundaries before actions reach the control plane
Make agent actions observable and auditable using existing cloud-native telemetry
Create interoperable agent ecosystems aligned with CNCF principles
The result is a shift from opaque, agent-driven automation to predictable, contract-based operations. By treating KMCP as a control plane interface, platform teams gain the same guarantees they expect from APIs, controllers, and CRDs - applied to AI-driven systems.
This session is intended for platform engineers, Kubernetes practitioners, and architects designing the next generation of AI-native infrastructure.
GPU‑Accelerated Workloads on KubeVirt: Scaling ML/AI in Kubernetes
KubeVirt is redefining how we run virtual machines in Kubernetes but what happens when those VMs need GPU acceleration for demanding AI/ML workloads?
In this lightning talk, I will walk through how to enable GPU-backed virtual machines in KubeVirt, and why this approach is gaining traction for secure, scalable, and isolated inference pipelines. We will explore the differences between container-based and VM-based GPU allocation, and see how KubeVirt integrates with CNCF tools like Prometheus and Kubernetes scheduler to monitor and optimize performance. If you are looking to push KubeVirt beyond typical VM use cases and into production-ready ML/AI workloads, this session will give you the technical foundation and the inspiration to get started.
Cloud Native Artificial Intelligence:Building Self-Healing Cloud Native Infrastructure with AI
This session explores how combining artificial intelligence and cloud native infrastructure can transform present Kubernetes operations into intelligent and self-managing platforms. Attendees will learn about how implementing AI-driven resource optimisation, predictive scaling, and automated incident response in cloud native environments will help organisations. This will equip machine learning engineers and data scientists with the knowledge to understand the changing Cloud Native Artificial Intelligence (CNAI) ecosystem and its opportunities.
TinyML at the Edge: Deploying and Optimizing AI Workloads on Zephyr RTOS
TinyML is transforming edge computing by enabling smart inference directly on microcontrollers but resource limitations make deployment complex. Zephyr RTOS lightweight, modular, and feature-rich is becoming a go-to platform for building embedded AI systems.
This session walks through how to effectively run TinyML workloads on Zephyr using various inference engines like TensorFlow Lite Micro, microTVM, emlearn, and LiteRT, along with decision points for selecting runtimes based on hardware constraints.
We will explore the runtime and how it simplifies AutoML workflows while supporting multiple backends. Attendees will also learn to use Zephyr’s Linkable Loadable Extensions (LLEXT) for hot-swapping models without reflashing. Performance optimization techniques such as quantization and operator fusion will be covered, along with benchmarking on physical devices vs Renode simulation.
The talk concludes with real-world examples like health monitors and predictive maintenance, best practices for OTA model updates, and the future of embedded AI with Zephyr.
KCD Sri Lanka 2025 Sessionize Event
KubeVirt Summit 2025 Sessionize Event
Aman Mundra
Founder & CEO, Welzin | Co-founder & CEO, CogNerd
Chandigarh, India
Actions
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top