ALD Virtual | Schedule

This session presents a practical case study of Instacart’s product substitution system, highlighting the real-world engineering tradeoffs behind deploying recommendation models at scale. Rather than focusing on novel algorithms, it examines challenges like cold-start handling, retailer-specific constraints, personalization vs. robustness, and when simpler models outperform complex ones in production. The talk also covers monitoring, iteration, and operational evolution, offering data scientists and ML engineers concrete lessons on building and maintaining reliable, large-scale recommendation systems under real-world constraints.

This session examines four recurring statistical pitfalls in industry experimentation that can distort high-stakes business decisions. Drawing on a decade of experience leading A/B tests and causal impact analyses, it highlights common but often overlooked errors that introduce bias or inflate variance in contexts such as pricing, promotions, and ad optimization. Through concrete examples, the talk explains the underlying statistical mechanisms, provides clear diagnostic approaches, and offers practical remedies for designing more reliable and defensible causal analyses.

As AI agents evolve into long-running, autonomous systems, designing effective long-term memory becomes essential. This session explores practical, production-ready memory architectures, including hierarchical and vector-based storage, multi-modal embeddings, and strategies for refresh, pruning, and consistency. It also addresses real-world challenges such as staleness, retrieval latency, and context explosion, with guidance on integrating memory into planning and control modules. Attendees will gain actionable patterns for building scalable memory systems that enable agents to reason, recall, and operate reliably in dynamic environments.

This session explores how large-scale A/B testing programs can evolve from isolated experiments into reusable knowledge systems. By systematically analyzing historical test repositories, organizations can identify patterns, avoid repeated pitfalls, replicate successful strategies, and improve the design of future experiments. The talk also examines how AI and autonomous agents can mine vast experimental datasets to enhance statistical precision, reduce required sample sizes, and generate insights beyond manual analysis—turning past experimentation into a compounding asset that accelerates innovation and decision-making.

This session presents a Multi-Agent System (MAS) framework for automating KPI root-cause analysis in complex, cloud-native enterprise environments. As distributed microservices, hybrid architectures, and dynamic workloads make manual triage increasingly difficult, the proposed architecture uses domain-specialized agents to monitor signals, correlate anomalies, and reason across temporal and causal relationships. By combining distributed decision-making, shared knowledge graphs, rule-based inference, and ML-driven anomaly detection, the approach reduces MTTD and MTTR while improving explainability—offering a scalable model for proactive reliability engineering in modern data platforms.

This session explores how enterprise AI decision-making practices can be adapted to safety-sensitive consumer products such as topical and ingestible wellness solutions. Using the Evidence-First Functional Formulation Design (EFFFD) approach, it focuses on structuring decisions around clearly defined problem boundaries, required evidence, and explicit guardrails rather than model performance alone. Through practical examples, the talk demonstrates how to evaluate uncertainty, manage trade-offs, and make defensible, evidence-bounded decisions in environments with limited data and real-world constraints—offering an actionable framework for extending AI rigor into consumer-facing contexts where trust and safety are critical.

This session examines how forecasting is shifting from static prediction engines to agentic, goal-driven systems that continuously adapt to changing conditions. Drawing on applied experience in the CPG industry, it explores how agent-based architectures extend strong statistical foundations to dynamically interact with pricing, supply chain, promotion, and operational constraints in volatile environments. The talk highlights how agentic forecasting differs from traditional predictive models, the importance of robustness and uncertainty management, and how multi-agent coordination can improve decision quality while preserving statistical rigor—offering a practical view of what works in production and where the real value lies.

This session presents a practical framework for model explainability and observability across the full ML lifecycle, combining local and global explanation techniques with system-level monitoring. Designed to support legal, compliance, and policy stakeholders alongside technical teams, the approach enables models to be interpreted, validated, and audited in production. It also supports continuous monitoring and drift detection at scale, significantly improving anomaly identification and reducing manual investigation effort—offering a blueprint for deploying transparent, accountable AI systems in enterprise environments.

This session explores how retail is shifting from AI-driven optimization to AI-driven decision-making in real time. Moving beyond traditional forecasting and recommendation systems, the next phase of retail technology centers on autonomous, self-learning systems that perceive, reason, and act across merchandising, supply chain, and customer experience. The talk examines how foundation models, conversational analytics, and human-in-the-loop design are enabling end-to-end AI systems that function as decision copilots—augmenting human judgment at scale rather than replacing it.

This session explores how to determine which e-commerce items are truly “speed-sensitive,” balancing customer satisfaction with margin control. Using large-scale historical sales data, it addresses the challenge of observational bias—where promotions, seasonality, and external factors distort demand signals. The talk outlines how causal inference techniques combined with deep learning can isolate the true lift generated by faster shipping, including approaches such as Inverse Propensity Treatment Weighting (IPTW), two-stage modeling architectures, and scalable backtesting across millions of SKUs. By emulating randomized controlled trials using observational data, the framework enables data-driven shipping decisions without costly live experimentation, helping optimize revenue and operational efficiency at scale.