Writing on how decisions get modeled, built, and deployed.
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.
Interactive tools and calculators for practitioners — built to explore ideas, not to sell software.
Available for consulting engagements in forecasting, pricing, and decision systems. More about me