HALL 3 - Tech Talks

As AI agents move from prototypes to real-world deployment, ensuring their reliability, transparency, and performance becomes critical. This session explores practical approaches to evaluating and debugging AI agents by tracing their decision-making processes, identifying failure points, and implementing robust testing strategies. It will cover methods for monitoring agent behavior, improving observability, and validating outcomes to build trustworthy and production-ready systems. Attendees will gain insights into tools, frameworks, and best practices that help developers and organizations confidently deploy and scale agentic AI solutions.

Physical intelligence, the ability of machines to perceive, reason, and act in the physical world remains one of the key frontiers in artificial intelligence. While modern AI systems have made significant progress in language, vision, and pattern recognition, they still struggle to fully understand and interact with the physical laws that govern real-world environments. This session explores the technology gaps in physical intelligence, focusing on the disconnect between current intelligence models and the principles of physics that shape real-world interactions. It will examine limitations such as weak physical reasoning, challenges in predicting object dynamics, and the difficulty of learning from limited real-world data. The talk will also highlight the “physics gaps” in today’s AI models, where purely data-driven approaches fall short in capturing causal and dynamic properties of the physical world and discuss emerging opportunities in areas such as embodied AI, simulation-based learning, robotics, and hybrid physics-AI systems that aim to bridge these gaps.

As large language models move from prototypes into enterprise workflows, teams across the industry increasingly face a practical operational alignment problem: how to steer model behavior toward business outcomes in a reliable and measurable way. At the same time for many practitioners a critical decision remains unclear: whether a use case is best addressed through prompt engineering, supervised fine-tuning, or reinforcement learning. This talk introduces a structured framework for choosing the right approach and a practical method for translating business objectives into reward signals that can be optimized, evaluated, and audited.
Through hands on experiments in incremental preference learning using reward modeling and policy optimization, the session demonstrates how meaningful behavioral shifts can be achieved even with small, carefully curated datasets. It also examines the stability trade-offs between staged training and online RL updates. Additionally, the session distills practical guidance on designing verifiable reward signals from ambiguous business objectives. This segment identifies when exploration and delayed rewards make reinforcement learning necessary and how to avoid common failure modes, such as reward hacking and instability. Rather than focusing on scale alone, the session emphasizes disciplined reward design and systematic experimentation as the foundation for deploying reinforcement learning effectively in enterprise LLM systems.

LLMs today promise endless text generation, but creating high-fidelity synthetic text data that actually reflects complex business logic remains an engineering challenge. This talk moves beyond basic “prompt and pray” techniques to address the various nuances of creating instruction datasets useful for knowledge distillation, domain adaptation, and reinforcement learning (RL) workflows.
We will examine why direct generation often causes datasets to regress to the mean, producing repetitive, safe content that lacks the messy edge cases required for robust training. To solve this, we suggest a systematic, algorithmic approach that treats data generation as an engineering problem. We will discuss how to decompose pipelines into iterative batches to programmatically inject real-world variations. We also conclude with a strategic checklist to evaluate if synthetic data is truly well-suited to your enterprise problem.

As AI agents become more prevalent across industries, ensuring they operate with the depth and accuracy of true subject matter expertise is becoming increasingly important. An expert-first approach focuses on anchoring AI agents in structured domain knowledge, expert insights, and reliable information sources rather than relying solely on generalized model outputs.
This session will explore why grounding AI agents in subject matter knowledge is essential for building trustworthy and effective AI systems. It will discuss the role of expert input, curated knowledge sources, and contextual understanding in enabling agents to deliver more accurate, relevant, and meaningful outcomes. The session will also highlight how organizations can design AI solutions that combine the power of large language models with domain expertise to create more reliable and impactful AI-driven experiences.

Every database architecture decision we have made relied on a few assumptions: callers are predictable, writes are deliberate, connections are short, failures are obvious, and schemas are understood. For decades, these held because a human was always in the loop.

Agentic AI changes that. Autonomous, LLM-powered agents generate queries through reasoning, write at machine speed, hold connections during long chains of thought, fail quietly, and interpret schemas through a model rather than shared context. When we attach them to existing data layers, the assumptions we never formalized start to show their cracks.

This talk walks through those assumptions, the production issues that follow, and how to design databases for agents. Database patterns that once seemed nice-to-have become essential once agents are in the system.

Agentic AI systems are rapidly moving from demos to mission‑critical workflows, but most of them still behave as pattern‑matchers with tools, not as systems that understand cause and effect. The result is familiar: agents that sound confident while suggesting actions that quietly violate business logic, break regulations, or create hidden risk. This talk introduces “causal guardrails”—an architecture where structural causal models (SCMs) sit around GenAI agents to constrain, explain, and validate their decisions. Instead of relying solely on prompts and heuristics, agents must route their plans through explicit causal graphs that encode allowed interventions, downstream impacts, and hard constraints. The session will walk through intuitive examples (credit risk, IT ops, or recommendation workflows), show how to combine LLM-based agents with SCMs in practice, and discuss how this improves robustness, debuggability, and auditability. Attendees will leave with concrete patterns for using causal modeling to keep autonomous GenAI “sane” in real enterprise environments, not just in benchmarks.

The Green Orchestrator proposes a next-generation agentic AI framework designed to coordinate, optimize, and govern distributed energy ecosystems operating at up to 1,000 TWh annual scale. As global energy systems become increasingly decentralized — spanning smart grids, renewable assets, data centers, EV infrastructure, and industrial facilities — existing optimization approaches remain fragmented, reactive, and limited to local objectives. Current AI deployments in energy largely function as advisory tools or isolated predictive models, lacking persistent memory, cross-system coordination, policy-aware autonomy, and multi-objective optimization capabilities.

This proposal introduces a hierarchical, multi-agent orchestration platform built using structured execution graphs (e.g., frameworks such as LangGraph), transforming large language models from conversational systems into goal-directed, stateful decision agents. Unlike conventional AI pipelines, the Green Orchestrator embeds agents within a deterministic, policy-constrained state machine architecture that supports long-horizon reasoning, controlled autonomy, and enterprise-grade observability.

At its core, the platform formalizes each agent as a constrained decision process operating over partially observable system states. Agents maintain belief representations through layered memory architectures consisting of short-term operational context, episodic summaries, and long-term vector-symbolic knowledge graphs. A novel energy-weighted memory optimization mechanism dynamically prioritizes retention based on carbon impact, financial risk exposure, grid stability sensitivity, and regulatory criticality. This approach significantly reduces token overhead while preserving high-value contextual intelligence, enabling scalable deployment across distributed edge environments. The system introduces hierarchical coordination across four layers: global strategic agents, regional grid agents, site-level optimization agents, and asset-level micro agents. Each layer operates within bounded authority while exchanging structured state updates. This creates distributed intelligence with escalation control and conflict resolution mechanisms analogous to enterprise governance structures. Multi-agent interaction is modeled as a stochastic cooperative game with weighted global objectives, enabling simultaneous optimization of energy efficiency, carbon reduction, cost management, resilience, and compliance.

A policy-bound autonomy framework ensures that all agent actions pass through validation gates including regulatory constraint checks, digital twin simulations, and risk evaluation layers before execution. This governance-first design differentiates the platform from experimental agent systems by embedding compliance and safety directly into the decision lifecycle. Domain knowledge is integrated through a hybrid approach combining pretrained model capabilities, retrieval-augmented access to enterprise documentation, structured ontologies of energy assets and constraints, and reinforcement learning via simulation environments. Agents leverage defined tool interfaces — including telemetry APIs, market data feeds, storage dispatch systems, and reporting engines — to interact with operational technology (OT) and enterprise systems in a controlled and auditable manner. The architecture is event-driven, activating agents only when triggered by system changes, thereby reducing computational overhead. Federated edge memory allows localized reasoning while sharing compressed embeddings upward, supporting data sovereignty and low-latency control.

Projected system impact at 1,000 TWh scale indicates that even modest coordinated optimization (8–12%) yields substantial reductions in energy consumption and carbon emissions while improving peak demand management and operational resilience. For enterprises such as Schneider Electric, the platform represents a strategic evolution from intelligent hardware integration to AI-native sustainability orchestration, enabling subscription-based optimization services and defensible intellectual property in policy-aware autonomous control.

In summary, the Green Orchestrator advances the field of agentic AI by integrating hierarchical multi-agent coordination, memory-efficient long-horizon reasoning, policy-embedded governance, and multi-objective optimization within a scalable enterprise framework. It establishes the foundation for a planetary-scale energy nervous system capable of learning, adapting, and autonomously coordinating distributed energy infrastructures responsibly and sustainably.

Agentic AI is reshaping how intelligent systems reason and act — but the next frontier lies in bringing that intelligence into the physical world. This session explores the shift from digital agents to Physical AI systems that interact with real-world environments, devices, and operations. We’ll examine the architectural principles, governance models, and system-level design patterns required to build reliable, scalable intelligent systems beyond the screen.

Most enterprise AI systems stop at generating predictions such as churn probabilities, fraud scores, recommendations, or forecasts, but business value is realized only when those predictions translate into reliable, automated decisions. This session focuses on the critical decision layer that sits between model outputs and real-world enterprise workflows. Designed for developers and ML practitioners, it explores how to build production-ready systems that combine model predictions with rules, thresholds, optimization logic, and human-in-the-loop controls to drive actionable outcomes. The talk will also cover handling uncertainty, edge cases, governance, and monitoring decision quality—not just model accuracy—ensuring AI systems are robust, scalable, and aligned with measurable business impact.