Agenda 2026

May 5-6, 2026 | The Clift Royal Sonesta, San Francisco

The most ideal way to soften the AI bubble’s looming detonation would be to boost AI’s realized value. How? A new reliability layer that tames large language models.

Enter predictive AI. LLM-based projects are usually too unreliable to move from pilot to production, but predictive AI can realize genAI’s bold, often overzealous promise of autonomy – or at least a great deal of it.

Join MLW conference founder Eric Siegel to learn why predictive AI is as crucial as ever in this “genAI world,” especially since it is positioned to solve genAI’s deadly reliability problem.

IBM Chief Data Scientist Kirk Mettler is one of those seasoned analytics veterans who don’t buy most of the AI hype – but still see a vital role for LLMs. His perspective plays out fortuitously as he dives into the weeds of advanced projects, even while he manages a large team of data scientists, data engineers, and other data professionals. “Today’s sky-rocketing investment into LLMs is just plain inappropriate,” Mettler says, “And yet, not leveraging LLMs to augment predictive AI projects would be just as inappropriate! We hit the sweet spot by blending LLMs with the time-honored tradition of enterprise machine learning.”

Conference Founder Eric Siegel will interview Mettler, digging in for what promises to be an extremely enlightening fireside keynote chat.

In the world of fraud detection, generative AI has unlocked remarkable new capabilities for analytics, and as a result, companies are racing to develop ways to explore data, summarize evidence, and build fraud-detection workflows. It has become clear that hybrid solutions – combining genAI and traditional machine learning –  are most effective when solid business use cases are paired with curated data, human oversight, and statistical rigor. 

Attendees will learn why fraud detection remains stubbornly difficult in retail and other verticals, how to design a genAI workflow with human-in-the-loop safeguards, and how ML models and genAI systems complement each other rather than compete to deliver real value for organizations. 

As enterprises rapidly adopt AI agents (e.g., Salesforce’s Agentforce), a critical risk emerges: misconfigured or compromised agents performing anomalous, potentially harmful, data operations. This presentation unveils an original, practical methodology for detecting such threats using unsupervised machine learning.

Drawing from a real-world Proof-of-Concept, Millie demonstrates how behavioral profiling—analyzing features engineered from system logs like data access patterns, query syntax (SOQL keyword analysis), and IP usage, along with signals from the content moderation mechanisms embedded within the LLM guardrails such prompt injection detection and toxicity scoring—can distinguish risky agent actions. Explore the creation of 30+ behavioral features and the application of KMeans clustering to identify agents exhibiting statistically significant deviations, serving as an early warning for misuse or overpermissive configurations. Millie will share insights into observed differences between AI agent and human user profiles, and challenges like crucial data gaps that impact comprehensive monitoring.

This session offers a vendor-neutral, technical deep-dive into a novel approach for safeguarding enterprise AI deployments.

Learning Objectives for Attendees:

1. Understand the novel security risks posed by misconfigured/overpermissive enterprise AI agents.
2. Learn a practical methodology for behavioral profiling of AI agents using unsupervised ML and log data.
3. Identify key data features, feature engineering techniques (e.g., for SOQL analysis), and common data challenges (log gaps, attribution) in AI agent monitoring.
4. Gain actionable insights to develop proactive detection strategies for anomalous AI agent activity and protect sensitive data.

Generative AI is redefining how enterprises design, deploy, and optimize data systems. This session explores a production-grade GenAI framework that automated data modeling, validation, and GCP infrastructure provisioning—reducing manual effort by 70% and accelerating release velocity by 40%. Attendees will learn how to design scalable GenAI pipelines, integrate LLMs with enterprise data workflows, and apply intelligent automation to streamline complex engineering tasks. Drawing from real-world deployment experience, this talk reveals actionable strategies for transforming AI innovation into tangible operational efficiency across large-scale data ecosystems.

Predictive AI offers tremendous potential – but it has a notoriously poor track record. Outside Big Tech and a handful of other leading companies, most initiatives fail to deploy, never realizing value. Why? Data professionals aren’t equipped to sell deployment to the business. The technical performance metrics they typically report on do not align with business goals – and mean nothing to decision makers.

What’s needed is an OTS solution that empowers stakeholders and data scientists alike to plan, sell, and greenlight predictive AI deployment. This serves to establish and maximize the value of each machine learning model in terms of business outcomes like profit, savings – or any KPI. 

Only by measuring value can the project actually pursue value. And only by getting business and data professionals onto the same value-oriented page can the initiative move forward and deploy.

In this session, Gooder AI CEO Eric Siegel will demonstrate a new category of software, ML valuation. The first such software solution, Gooder AI can valuate any ML model, effortlessly serving up an interactive visualization to navigate ML deployment and make the value easily understood by business stakeholders.

Information will follow shortly.

GenAI pilots often stall at the last mile—reliability. HP ships a hybrid AI reliability layer that fuses generative AI with predictive ML and humans-in-the-loop to make deployments robust. Samaresh shares the production patterns HP use across HP platforms (e.g., Z Boost, AI Studio, Hydra): guardrail policies that expand over time, retrieval and output validation with predictive models, confidence-aware routing between GenAI/predictive/HITL, and telemetry (success rate, hallucination/override rate, p95 latency, cost). Attendees get a reference architecture, rollout playbook (canary/shadow), and KPI templates to move from demo to dependable—turning AI’s potential into realized value.

More information about this keynote address is coming soon — about Anori, X’s moonshot to make building and development faster, more affordable and more efficient.

Information will follow shortly.

Decision agents drive autonomy in AI systems. They combine generative AI or large language models for interaction and explanation, with machine learning for precision, and use decision models and business rules to act as a reliability layer. They allow each technology to play the role its best suited for. And, by focusing on the business decision being made rather than the specific technology being used, they keep business owners engaged.

This session will use real-world examples in banking, healthcare and insurance to illustrate the power of decision agents to personalize interactions, operationalize machine learning and augment human decision-makers. It will highlight how decision agents maximize the power of each technology, outline a proven development approach for decision agents, and share some dos and don’ts based on years of experience in developing autonomous systems.

The ‘magic moment’ for AI agents is hidden in the memory layer. Context Engineering is the art of shaping what an AI model knows at any moment by managing how information enters, persists, or fades from its working memory. In this session, we explore how machine learning systems “remember” through state objects, notes, and retrieval—and how they manage context using compression, selection, and context limits. We’ll walk through real-world agent patterns that balance personalization with privacy, performance, and relevance. Participants will learn practical techniques to design memory that feels intentional, evolving, and human-aware.

Come learn about how Twilio’s AI assistant prototype tripled customer conversions and scaled to support 80 countries and 12 languages. Rikki will discuss real application of ML within Twilio’s functional environments to accelerate growth.

Twilio’s AI assistant inspired this article on genAI’s reliability layer by conference chair Eric Siegel: https://www.forbes.com/sites/e...

Information will follow shortly.

Suvendu will present a novel service-oriented architecture for scaling Reinforcement Learning from Human Feedback (RLHF) across heterogeneous accelerators, specifically targeting the migration from GPUs to AWS Trainium. Our approach addresses key challenges in implementing complex RLHF pipelines that orchestrate multiple models (SFT, Actor, Critic, and Reward Model) while supporting models up to 1T parameters for dense architectures and 2T parameters for Mixture-of-Experts (MoE).

The proposed system — which we have successfully applied this approach for for a general-purpose conversational AI product — introduces three key innovations: (1) A microservices-based architecture that separates non-Actor components across nodes, enabling flexible scaling and efficient resource utilization across GPU and Trainium accelerators; (2) A novel pipelined generation method that overlaps Actor inference and training using a consumer-producer buffer, significantly reducing training latency; and (3) Support for advanced techniques including LoRA fine-tuning, Grouped Query Attention, and multiple reward models.

This implementation leverages Neuron Nemo Megatron for cross-platform compatibility while incorporating optimizations for both GPU and Trainium backends. The architecture enables efficient handling of long contexts (>200k tokens) and provides a pathway for future optimizations including slop buffer implementation and lightweight multi-tenant reward modeling.

Learn proven methodologies for deploying reliable generative AI agents in production using serverless architecture. This presentation covers essential patterns for building stateless, user-aware LLM systems that scale automatically while maintaining security and reliability. We’ll explore Model Context Protocol (MCP) for secure agent-to-tool communication, JWT-based authentication flows, session state externalization strategies, and reliability layers that tame LLM unpredictability. Attendees will discover practical techniques for overcoming common genAI deployment challenges including state management, user context propagation, tool integration, and production readiness. The methodology demonstrates how to bridge the gap between genAI pilots and production systems through hybrid architectural approaches.

This session explores the design and deployment of enterprise-scale anomaly detection platforms serving Finance, Audit, and Compliance functions. Enterprise governance demands AI systems that balance risk detection, operational efficiency, and regulatory compliance. This talk explores building production-grade anomaly detection platforms serving Finance, Audit, and Compliance. Beginning with the evolution of governance AI adoption, Vineeth will discuss Schneider Electric’s Anomaly Detection Framework encompassing metric development (KPIs), feedback-as-a-service architecture, continuous model monitoring and updates, fallback model strategies for system resilience, and policy adherence mechanisms. We’ll examine MLOps infrastructure(AWS) for cross-functional deployment, interpretability requirements in high-stakes decisions, and balancing detection sensitivity with false positive rates. Engineering details highlight AWS-based  model inference, retraining, model versioning strategies, and patterns for scaling anomaly detection across organizational silos while maintaining regulatory compliance.

Henry Castellanos has lived in the analytics trenches across a wide range of enterprises. He believes that realizing value with predictive AI is almost always an extreme challenge that’s met only with special attention to robustness, scale, and explicit KPIs.

In recent years, Henry has found that genAI serves a crucial role in this. Rather than usurping analytics projects — which have only grown more important — genAI empowers practitioners to develop a new degree of robustness and validation for all things quantitative, from anomaly detection to predictive AI.

Come to Henry’s keynote address to hear his unparalleled story and lessons-learned.

Modern generative AI systems—from LLMs to multimodal models—are no longer compute-bound; they are memory-bound. As model sizes soar, inference latency is dominated by memory bandwidth, memory fragmentation, KV-cache bloat, checkpoint restore time, and PCIe/NVLink bottlenecks. This session breaks down the “Memory Wall” limiting generative model performance and shares practical techniques such as model compression, quantization, memory-efficient attention, sharding, and cold-start optimization. This talk provides actionable insights for practitioners building large-scale generative AI infrastructure.

Generative AI is not only great at generating artificial content, it is great at personalizing and recommending human-created content: from Music, to Videos, to Podcasts, to Audiobooks. Through the use of Large Language Models, we better understand the content creators created and the content consumers are interested in. This allows GenAI to find the best content for each user as well as give context on why this new content is relevant. Context can come from an AI DJ explanation or the right image or preview or trailer. By integrating world knowledge, content understanding and user understanding, LLMs provide a unique opportunity to craft what we call personal narratives—stories that resonate with listeners and familiarize them with recommendations.

In this session, Gooder AI CTO Luba Gloukhova will cover the most underutilized – yet shockingly simple – way to multiply predictive AI’s value. Instead of driving decisions with the raw model score, drive decisions with the EXPECTED VALUE (calculated with the model score, and at least one business-side factor). This represents a paradigm shift only because it is very rare and very impactful – it’s a straightforward “no-brainer,” both technically and conceptually.

Luba will show demos of how this work – and how incredibly WELL it works – across several use cases.

Fraud detection systems often face unreliable ground truth due to costly and inconsistent human labeling. This session presents agent-in-the-loop framework, which uses large language models (LLMs) as intelligent labelers to power a continuous learning flywheel. The LLM integrates domain knowledge and historical data to generate high-quality labels, propose new predictive signals, and enhance model retraining. Through agent-in-the-loop feedback and quality controls, the system continually refines its understanding. Attendees will learn how adaptive LLM-ML systems can drive scalable and self-improving fraud detection models.

As organizations scale their use of machine learning and generative AI, the cost of model failure grows. CSAA Insurance Group built an Independent Model Validation (IMV) team to evaluate models across multiple risk dimensions: bias, fairness, robustness, explainability, and real-world impact. In this session, Aaron will share how CSAA operationalized scalable model validation across supervised models, large language models, and vendor-built systems. You’ll hear real examples of model flaws they’ve caught and how they balance rigor with innovation. Model validation isn’t a gate, it’s how machine learning becomes safe, fair, and launch-ready.

With an extensive record serving executive roles at the likes of Starbucks, GM, and State Farm, Jon Francis has learned what makes the difference between scalable AI systems and those that never get beyond pilots or even just ideation. “Here’s a hint,” he says. “At most companies, it has almost nothing to do with the math or technology, but instead hinges on culture, transformation, and change management.”

Join this fireside-chat keynote, where Machine Learning Week founder Eric Siegel will interview Francis and dive in.

How much of AI’s presumed value and its promise of supreme autonomy is real, and how much is hype? Is it 20% hype? 60% hype?

Our expert panelists have some thoughts. Join this session as they pinpoint AI’s lowest hanging fruits, navigate how to squeeze proven value out of AI’s sometimes-audacious value propositions, and carefully maneuver the management of that incorrigible breed of AI hype that overpromises.

High volume transaction systems are core to many large enterprises. Incorporating scoring (inferencing) based on traditional structured data values within these transaction systems could be challenging. Adding dense, rich text analysis through generative AI adds yet another dimension of complexity. This session will provide an overview of the alternative approaches to incorporating scoring within transactional systems, review the pros and cons and share the way that IBM is working with our customers to incorporate real-time scoring within their transactional systems.

The AI ecosystem is flooded with demos that fail in production due to cost, scalability, or reliability gaps. This challenge is amplified in FMCG, where millions of interconnected outlets, SKUs, and supply nodes offer opportunities for AI enablement at scale. At Bizom, India’s top Supply chain automation provider, we’ve engineered a hybrid AI framework combining generative and predictive methods. Generative models create contextual training data, while lightweight predictive models—like XGBoost—enable scalable, low-cost deployment. This approach transforms fragile prototypes into production-grade intelligence for the FMCG sector. In this session, I will share field-tested use cases and technical insights on designing hybrid AI systems that turn the promise of AI into measurable business performance at scale.

As AI adoption accelerates, organizations must ensure their generative and predictive models remain reliable and trustworthy in production. This session explores how to embed full-stack telemetry across data, model, and infrastructure layers to detect issues such as data and concept drift, latency degradation, distribution shifts, and hallucination patterns in generative systems. Participants will learn practical techniques for building monitoring and alerting pipelines, establishing feedback loops for retraining and governance, and integrating human oversight where needed. The discussion also highlights the organizational practices required to operate hybrid AI systems effectively in real-world environments, using concrete examples from ML pipelines and modern observability platforms.

This demo walks through a hands-on AI effort where big language models and classic ML methods tried guessing Amazon product costs using heaps of customer reviews. Starting off with a 5GB batch of 46,726 home gadgets priced between $0 and $100, the scope later grew to cover car parts, gadgets, office supplies, hardware, phones, and toys – pulling in titles, descriptions, specs, and extra details. Instead of stacking models together, each one was tested solo: linear regression, SVR, and random forest – fed with custom features, bag-of-words, plus word2vec encodings. Results swung from a shaky $139 off using plain regression down to just $46 error – alongside an RMSLE of 0.38 – thanks to sharp-tuned, slimmed-down LLaMA 8B setups and other cutting-edge designs. All training and checks were tracked live on wandb.ai across 400K samples, revealing that deeper training runs plus tight tuning push accuracy way up.

Timely identification of rare disease patients remains a significant challenge, as early phenotypic clues are often embedded in unstructured physician documentation across multiple specialties. Shantanu presents a multi-model large language framework that leverages a large-scale corpus of de-identified physician notes—representing over 300 million patient lives—to uncover potentially undiagnosed rare disease patients through layered text analytics and AI-driven reasoning.

Axtria’s approach combines deterministic information retrieval with contextual inference across three complementary LLMs — GPT-5, GPT-5 Mini, and Sonnet 4 — orchestrated in a sequential discovery pipeline:

1. Targeted String Search: They first execute high-precision string and semantic pattern searches to isolate note fragments containing disease-relevant terms, symptom mentions, and phenotype descriptors.

2. Patient Linking: Each relevant note is mapped back to the originating patient to establish a cohort of individuals with preliminary indicators of the target disease.

3. Comprehensive Contextual Review: For these patients of interest, all available clinical notes—spanning primary, specialty, and ancillary care—are aggregated to capture longitudinal context and diagnostic evolution.

4. Cross-Domain Triangulation: Using GPT-5 for deep contextual reasoning and Sonnet 4 for meta-analysis, the framework triangulates laboratory results, medication histories, symptom progressions, and physician impressions to identify patients exhibiting consistent disease signatures despite lacking explicit diagnostic codes.

This multi-model ensemble produces interpretable outputs with traceable rationale chains, enhancing clinician trust and enabling focused chart review. Across pilot rare-disease cohorts, the framework achieved a 4–5× improvement in patient identification rates versus ICD-based baselines, with physician validation confirming >80% contextual accuracy.

By fusing deterministic filtering with hierarchical LLM reasoning, this framework demonstrates a scalable, privacy-preserving approach for surfacing undiagnosed patients, accelerating early intervention, and informing precision outreach strategies. Future extensions will integrate claims, genomic, and laboratory data to further strengthen predictive fidelity and clinical utility.

The real meaning to Machine Learning algorithms are its real world impacts. There could be many different ways in which the customer experience can be enhanced. Businesses can analyze the customer preferences and history to tailor their services. Machine Learning algorithms also help in reducing the customer support time and curate smart checkins, checkouts as well as smart ordering, which in turn increases customer satisfaction. Basic customer enquiries could be handled by ML chatbots. Demand forecasting aids in improved availability of rooms for guests. Fostering a brand’s loyalty and providing a seamless stay to customers is what leads to a 5 star service in the hospitality industry.