Writing
Deep technical analysis and strategic insights across data science, analytics, and business intelligence.
Latest Articles
53 results
Seven ML products, one retail team, and the question of whether to consolidate training into a single Vertex-native monorepo. The answer is conditional, and the condition is the whole point.
Why data science and AI platform projects fail on definitions, not on infrastructure, and how to let a real decision pull the data model into existence instead of waiting for the lake.
A staff-level field guide to predicted LTV for marketing activation: why ranking beats accuracy, why a directional model can beat an accurate one for seeding audiences, and where the whole thing quietly goes wrong.
Enterprise AI systems often work at the first level of granularity, then become fragile when the business asks for more precision. A field lesson from pharmaceutical supply optimization on why incremental architectures matter.
Raw latency and perceived latency are different engineering problems. Production GenAI systems feel fast when they expose progress early, overlap backend work, and avoid silent waiting.
What it actually takes to build a production forecasting system for luxury demand at SKU level: decomposing products into shared attributes, designing hierarchical aggregate features without leakage, handling intermittency with Tweedie and hurdle objectives, and forecasting in a way finance can use.
Intelligence-native systems need agent access to decision artefacts and feedback loops. Why context, not models, is the differentiator — and how MCP, traditional ML, and versioned artefacts fit together.
A field perspective on hierarchical forecasting in luxury retail — where ultra-premium SKUs sell fewer than 100 units a year and finance cannot ignore them.
Markdown planning explodes when weeks, events, discount levels, and phases are modeled as separate dimensions. Collapsing each week into a single state turns it into a clean optimization problem.
Seasons, campaigns, and weekly events — retail runs on overlapping cycles, and the AI recommender has to keep up with all of them. Notes on the business-rules control surface that lets merchandising teams steer a conversational recommender without editing prompts, filing tickets, or waiting for a deploy.
Macy's annonce que les clients utilisant son assistant IA dépensent 4,75x plus. Sephora vient de lancer une app dans ChatGPT. Zalando déploie son assistant dans 25 marchés. La question pour tous les autres retailers n'est plus 'faut-il le faire ?' mais 'comment l'architecturer pour que ça tienne en production ?'
This is the engineering companion to the production architecture piece. Instead of re-arguing why open-ended agents are risky in commerce, it walks through the actual implementation choices in `ai-florist`: FastAPI boundaries, LangGraph orchestration, pgvector retrieval, learned scoring weights, deterministic fallbacks, and runtime observability.
Macy's reports customers spending 4.75x more with their AI assistant. Sephora just launched inside ChatGPT. Zalando rolled out to 25 markets. The question for every other retailer isn't whether to build this — it's how to architect it so it actually works. Here's what I learned building one for premium floral e-commerce.
LangChain, LangGraph, Claude ADK, Google ADK, PydanticAI — different names, different abstractions, same promise. This article cuts through the framework noise and identifies the five architectural properties that actually determine whether an LLM system survives production.
LLMs can extract structured data from anything — until they cannot. This article documents the failure modes of LLM-only parsing in production pipelines, and presents a layered architecture where determinism comes first and the LLM is used only where it is structurally irreplaceable.
La plupart des échecs GenAI en entreprise ne viennent pas d'un mauvais modèle. Ils viennent d'un mauvais design de workflow, d'une absence de mandat et d'une architecture incapable de faire agir un système.
Most companies are not ready for autonomous commerce agents. The practical starting point is Agentic RAG: systems that retrieve business context, reason over it, and produce decision-ready outputs.
Two initiatives can show the same ROI and still deserve very different decisions. The missing variable is time-to-ROI: how fast the initial investment is recovered.
Sales volume mixes demand and inventory effects. Sell-through rate captures what actually matters: how fast inventory clears under pricing and context constraints.
Most CRM organizations don't have a feedback loop. They have a broadcasting system. The shift from propensity to uplift modeling changes everything.
Modern LLMs are objectively better at nuanced language tasks than legacy systems. But when you zoom out from a single prompt in a playground to a production environment processing millions of requests, the paradigm shifts entirely. Here is why using an LLM as a universal parser and router is often the wrong architectural choice.
Pourquoi le forecast de lancement produit est souvent un exercice politique plus qu'une estimation robuste, et comment le transformer en système probabiliste et explicable.
Revenue is a derived quantity. Marketing influences demand, units sold, not price. Modeling revenue directly entangles two mechanisms and produces misleading attribution.
A comprehensive guide to deploying production-ready Marketing Mix Models—covering data ingestion, model training, cloud deployment, and real-time optimization APIs for budget allocation
Deep dive into gradient boosting trees with quantile regression for robust uncertainty estimation in travel retail forecasting.
Learn how Mixed Integer Programming (MIP) with discretized response curves solves marketing budget allocation problems using branch-and-bound methods.
How to build robust forecasting models that account for uncertainty, seasonality, and external shocks in the travel retail sector.
Walk through real-world MMM optimization scenarios using Mixed Integer Programming. See how to discretize response curves, formulate constraints, and solve practical reallocation problems with actual numbers.
Exploring structural causal models using DAGs, counterfactual reasoning, and frameworks like DoWhy and PyMC-Marketing
How Google's Meridian brings causal inference to marketing mix modeling at global scale
Understanding why modern marketing optimization focuses on efficiency, not volume. Learn how marginal ROI transforms budget allocation and why the slope matters more than the height.
Understanding the mathematical law that explains why large brands not only have more buyers but also enjoy higher loyalty—and how it reshapes marketing strategy
Discover how the Michaelis-Menten equation models immediate diminishing returns in marketing. Learn when to use it instead of Hill functions and why it's the standard for direct response channels.
Understanding how MMM optimizers reallocate marketing spend across channels to maximize efficiency. Learn the step-by-step process from response curves to optimal budget allocation.
Master the economics of marketing spend through the five zones of efficiency, from minimum threshold to market saturation. Learn when to scale, optimize, or reallocate budget for maximum ROI.
Why retailers can't escape markdown pricing—a deep dive into the lifecycle, inventory, and pricing constraints that shape modern retail optimization strategies
Understanding why retail revenue functions are concave, how this property enables convex optimization, and the network flow interpretation that powers modern markdown pricing algorithms
A comprehensive guide to saturation curves and their application in media optimization. Learn why Hill functions can be concave or convex and how they're used in Google Meridian.
How retailers transform predictive demand models into prescriptive pricing decisions through stochastic programming, jointly optimizing markdown prices and inventory allocation under uncertainty
A deep dive into why measuring price elasticity remains the biggest challenge in demand modeling, from data scarcity and endogeneity to seasonal confounding—and how modern techniques address these challenges
How hierarchical modeling solves the challenge of sparse data in retail by pooling statistical strength across products, enabling reliable demand parameter estimation even for low-volume SKUs
Exploring the mathematical foundations and practical applications of different response curves in marketing effectiveness measurement.
How predictive demand models serve as the critical bridge between historical sales data and prescriptive markdown optimization, providing both demand estimates and uncertainty quantification
Understanding why prescriptive pricing models must respect the law of demand, and how this fundamental constraint shapes our choice between GLMs and machine learning approaches
Understanding the critical distinctions between causal identification, uplift, and contribution—and why conflating them leads to bad marketing decisions
A comprehensive introduction to Marketing Mix Modeling (MMM) — what it is, how it works, when to use it, and how it differs from attribution. Written for practitioners and decision-makers.
A structured breakdown of deploying an MMM solution, covering project initiation, data foundation, modeling, and operationalization based on real-world implementation patterns.
Analyzing 150 years of NYC's fish market data to understand how weather shocks reveal true price elasticity and market dynamics using causal inference
A critical analysis of the market share-profitability relationship through Bourantas & Mandes' dynamic model, revealing when bigger isn't always better
Explore geometric adstock decay with our interactive calculator. Model how your marketing spend creates lasting impact through carryover effects.
A practical guide to using conjoint analysis for understanding product preferences and pricing optimization
Using Bayesian demand modeling and mixed-integer programming to set optimal prices for a new product line — balancing revenue, cannibalization, and market share
Designing an automated credit decisioning system with gradient boosted trees — from manual review to real-time scoring in production