ALD SF | Schedule

On a whiteboard, an AI agent is simple: Prompt + Tools + Model. In production it explodes into twenty terms of infrastructure: sessions, streaming, memory, auth, tracing, orchestration. Most teams build agents the way we built monoliths fifteen years ago, each a self-contained stack that cannot talk to the next.

This talk applies two decades of microservices wisdom to agentic AI. The unit of value is a MicroAgent: an MCP server plus a prompt template, owned by one team, plugged into a shared platform that handles everything else. We cover MCP as the API contract, why memory should be shared across all agents, multi-agent composition, and the engineering economics of building infra once instead of N times. You leave with a concrete blueprint, not a framework survey.

Most enterprise AI agents fail not because of bad reasoning but because they cannot reliably understand what customers say. This session breaks down what actually breaks in production voice AI such as the accuracy gaps, language blind spots, and speed thresholds that determine whether an AI agent works or fails in the real world. Attendees will get a clear framework for evaluating the speech layer in their AI stack and the right questions to ask before shipping.

Every enterprise has AI pilots. Far fewer have enterprise-wide adoption.

Drawing on experiences at Snapchat, Meta, and PepsiCo, Daniela shares what separates successful AI initiatives from the ones that never move beyond experimentation. Through real-world case studies, she’ll explore how organizations identify high-impact use cases, overcome organizational resistance, measure business value, and build trust in AI across teams.

Attendees will leave with a practical framework for moving AI from proof of concept to measurable business impact, along with actionable lessons that leaders, product teams, and practitioners can apply within their own organizations.

As Agentic AI transforms customer-facing experiences across e-commerce, the difference between shipping with confidence and shipping blind often comes down to one capability: evaluations. This session introduces a practical framework for the continuous evaluation loop that powers production-ready Agentic AI experiences, combining offline evaluations that validate readiness before launch with online evaluations that measure real-world performance in production. We demonstrate how these complementary approaches form a unified feedback system for improving quality, safety, and business impact. Attendees will leave with a practical understanding of where evaluations fit within the agent lifecycle, how to design effective launch gates, and how to evolve from requirement-based delivery to hypothesis-driven iteration- without slowing innovation.

The convergence of AI and data streaming is redefining operational resilience. As modern systems become increasingly distributed and dynamic, enterprises need more than monitoring, they need infrastructure that can self-diagnose, self-correct, and continuously adapt. This session explores how real-time streaming pipelines can power agentic AI systems that deliver autonomous remediation at scale.

Drawing from deployments in telecom, manufacturing, and healthcare environments, we’ll unpack:

• How streaming telemetry fuels real-time inference, decision-making, and self-healing
• Key use cases where agentic AI delivered measurable uptime gains and OPEX reductions
• Design patterns for embedding explainability, guardrails, and human oversight into autonomous agents

Attendees will leave with an actionable blueprint for building intelligent, real-time infrastructure that moves from detection to correction in milliseconds, delivering trusted autonomy, resilience, and business continuity.

Modern AI systems are only as strong as the data infrastructure behind them. As organizations adopt machine learning, generative AI, analytics automation, and real-time decision systems, data teams must build platforms that are not only scalable, but also reliable, governed, observable, and cost-efficient.

This talk focuses on practical engineering patterns for building cost-efficient, AI-ready data infrastructure at scale. Drawing from my experience in fintech data engineering, I will discuss how teams can optimize SQL workloads, design efficient batch and incremental pipelines, build reusable curated datasets, implement data quality checks, and improve observability across production data systems.

The session will cover common causes of high cloud data costs, including unnecessary full refreshes, duplicated transformations, inefficient joins, poor data modeling, weak monitoring, and fragmented downstream usage. I will also explain how better data architecture supports AI readiness by improving freshness, lineage, trust, governance, and usability of data for analytics, machine learning, and agentic AI workflows.

Attendees will leave with a practical framework for identifying expensive or fragile data workflows and improving the reliability, scalability, and cost efficiency of their data platforms.

Enterprise adoption of Generative AI has largely started with Retrieval-Augmented Generation (RAG) systems that enhance large language models with organizational knowledge. While RAG has proven effective for search, question answering, and knowledge discovery, many enterprise use cases require more advanced capabilities such as planning, reasoning, workflow orchestration, tool usage, and autonomous decision support. This evolution is driving the emergence of Agentic AI and multi-agent architectures.

In this session, I will explore the architectural journey from traditional RAG implementations to enterprise-grade multi-agent systems. The talk will examine key design patterns for agent orchestration, memory management, context sharing, tool integration, and inter-agent communication using emerging standards such as the Model Context Protocol (MCP). Attendees will learn how specialized agents can collaborate to solve complex business problems that are difficult to address with a single AI model.

Drawing on experience building large-scale distributed systems and enterprise platforms, I will discuss practical considerations including scalability, governance, observability, security, and operational challenges associated with deploying intelligent agents in production environments. The session will also include architectural examples and a demonstration of how agent-based systems can be applied to real-world enterprise scenarios.

Attendees will leave with a practical blueprint for designing, evaluating, and implementing multi-agent AI architectures that move beyond traditional RAG and enable the next generation of intelligent enterprise applications.

AI analytics fails in production when teams treat reliability as a model problem instead of a system problem. A BI copilot or analytics agent can have a strong model, good prompts, and access to the right data, yet still produce the wrong KPI, ignore permissions, lose business context, or give an answer no one can audit.

This talk introduces the AI Analytics Reliability Stack: a practical framework for making GenAI analytics systems safer, more accurate, and production-ready. I will walk through five layers teams need beyond a basic chatbot: guardrails to control unsafe or out-of-policy behavior, grounding to connect answers to trusted enterprise data, evals to measure answer quality before and after release, lineage to trace where an answer came from, and human feedback to continuously improve the system.

Attendees will learn how these layers work together in real analytics workflows, including natural-language-to-SQL, RAG over business metrics, BI copilots, and agentic data tools. The goal is to help data and AI teams move from impressive demos to systems business users can actually trust.

Enterprise agentic AI is failing at scale — not because the models are wrong, but because the organizations are unprepared. Gartner forecasts over 40% of agentic projects will be cancelled by 2027. The root cause is consistent: brittle workflows, absent governance, and operators who were never set up to oversee autonomous systems.

This session introduces a two-stage framework for closing that gap. Stage 1 is a five-dimension readiness diagnostic that determines whether a workflow is actually ready for an agent before any model investment is made. Stage 2 is the agentic surface design — built around a single principle: human intelligence belongs where the agent is weakest and where the organization is most accountable.
Grounded in a Fortune 500 production deployment and peer-reviewed research, this talk gives practitioners a diagnostic they can run this quarter and a design they can hand to their team tomorrow.

Abstract coming soon!