Skip to content
Claudio Cabrera
Sequential Decision Analytics · Stochastic ML Systems

Mathematical decisions that compound margin.

Most e-commerce and supply chain operations bleed millions in myopic routing, stockout censoring, and uncalibrated safety buffers. I engineer production-grade, auditable decision engines grounded in Warren Powell’s Universal Modeling Framework—turning stochastic uncertainty into provable GAAP contribution margin.

Explore Live Systems →Run P&L Margin Simulator

UMF 5-Tuple

PFA · CFA · VFA · DLA Policies

100% Strict

Hexagonal Python · Type-Safe

Institutional

Chicago Booth · EY · BBVA

Canonical Decision Loop · Powell §1.8Sₜ → xₜ → Wₜ₊₁ → Sₜ₊₁

1. State Vector Sₜ = (Rₜ, Iₜ, Bₜ)

Physical inventory & biomass · SLA constraints · SMC particle belief over censored demand.

2. Feasible Action xₜ ∈ Xₜ

Multi-echelon fulfillment routing, split-shipment suppression, & dynamic B2B/B2C quote pricing.

3. Exogenous Shock Wₜ₊₁ & Physics Sᴹ

Spatio-temporal demand innovations, carrier transit delays, & immutable event-sourced transitions.

4. Objective C(Sₜ, xₜ, Wₜ₊₁)

Maximize Cumulative GAAP ASC 606 Net Contribution Margin subject to p95 Stockout SLAs.

Sequential Decision Science · Dynamic Pricing · Stochastic Supply Chains ·Sequential Decision Science · Dynamic Pricing · Stochastic Supply Chains ·

Interactive Proof-of-Work

Production Decision Systems

View Full Systems Catalogue →
Project Aegis: Autonomous Stochastic Supply Chain Orchestrator
in progress6 Interactive Modules
Powell's Sequential Decision FrameworkSpatio-Temporal Econometrics (Wikle)Finite Population Sampling (Lohr)Bayesian Experimental Design (Marzouk)Contribution Margin Optimization (Phillips)

Project Aegis: Autonomous Stochastic Supply Chain Orchestrator

An enterprise multi-echelon inventory allocation, dynamic pricing, and spatial fulfillment routing engine calibrated on 100,000+ Brazilian Olist e-commerce orders.

1.55M Records

Empirical Spine

K=5 Geodesic Hubs

Fulfillment Topology

p95 Stockout ≤ 3.5%

Primary Risk SLA

PFA · CFA · VFA · DLA

Policy Tournament

Python 3.14DuckDB OLAPHexagonal DDDPyTorch ST-GCNHypothesisDistributed SagasSequential Monte Carlo
Inspect Architecture & Modules →
Poultry CRM Operations System
complete6 Interactive Modules
Powell's Sequential Decision FrameworkDirect Lookahead Approximation (DLA)Analytical CRM · CLV & Churn ModelingRevenue Management (Talluri)Contribution Margin Optimization (Phillips)

Poultry CRM Operations System

A production-grade Direct Lookahead (DLA) pricing engine for B2B poultry wholesale — built on Powell's Sequential Decision Framework, deployed on GCP Cloud Run.

Class 4 DLA

Policy Class

9-Tier Rollout

Action Space

Atomic Outbox

Persistence SLA

Zero-Static OIDC

Trust Model

Python 3.14FastAPISQLAlchemyPrometheusDockerGCP Cloud RunFirebase Auth
Inspect Architecture & Modules →
Executive ROI Calculator

Quantify the cost of greedy heuristics.

Replacing myopic Nearest-Neighbor (PFA) rules with BOED-tuned Cost Function Approximations (CFA), Shadow-Price VFAs, or Direct Lookahead (DLA) policies directly recovers margin lost to split shipments, stockout abandonment, and poorly positioned safety stock.

Book a Counterfactual Margin Audit →

Counterfactual P&L Impact Estimator

GAAP ASC 606 Basis
$50M USD
$10M$125M$250M+
6.5%
2.0% (Lean)8.0% (Volatile)15.0% (Stressed)
5 Hubs
2 Hubs10 Hubs20 Hubs

Projected Margin Lift

+$1.62M

Annual Net Contribution

Stockout Reduction

6.5% → 2.8%

p95 Service Level SLA

Working Capital Freed

$1.15M

Via BOED/VFA Buffer Tuning

Core Capabilities

Where Theory Meets Production SLA.

Academic Operations Research models often fail in production due to mutable state bugs, lookahead leakage, or untyped glue code. Every system I architect enforces strict mathematical and software invariants from day one.

Explore Consulting Offerings →
01 · Warren Powell’s Universal Modeling Framework+

Complete mathematical separation of Physical State (Rₜ), Information State (Iₜ), and Epistemic Belief State (Bₜ), benchmarked systematically across all four policy meta-classes (PFA, CFA, VFA, DLA) against a formal Metric Pyramid.

02 · Dynamic Pricing & Censored Demand Probing+

Talluri & van Ryzin revenue management combined with Robert Phillips contribution margin optimization, active learning for stockout-censored demand, and multivariate cross-category covariance modeling.

03 · Hexagonal Python 3.14 & Distributed Sagas+

Ports & Adapters isolation, Alexis King’s "Parse, Don't Validate" ingestion gates, DuckDB analytical spines, transactional outboxes, and property-based Hypothesis tests proving state invariants across 10⁵ stochastic transitions.

04 · Counterfactual Off-Policy Evaluation & MLOps+

Seeded bit-for-bit replay determinism, Wasserstein distribution drift detection, zero-static-credential Workload Identity Federation, and shadow-serving tournament benchmarks before live traffic cutover.