Project Aegis: Autonomous Stochastic Supply Chain Orchestrator

Executive Summary

Modern multi-echelon e-commerce networks lose millions of dollars annually to three compounding pathologies: myopic nearest-warehouse routing (which prematurely depletes regional hubs and forces expensive long-zone split shipments), stockout-censored demand blindness, and uncalibrated safety stock buffers.

Project Aegis is an enterprise-grade autonomous sequential decision orchestrator grounded in 1.55 million empirical records (100,000+ customer orders) from the Brazilian Olist e-commerce ecosystem. Built on Warren Powell’s Universal Modeling Framework (UMF) and a strict Hexagonal (Ports & Adapters) Python architecture, Aegis benchmarks all four canonical policy meta-classes directly against GAAP ASC 606 Net Contribution Margin.

Architectural & Mathematical Foundations

1. Alexis King’s “Parse, Don’t Validate” Perimeter & DuckDB OLAP Spine: All 9 raw Olist relational tables (1,550,922 records) are validated at the ingestion boundary into frozen domain dataclasses and materialized into a zero-latency 76.76 MB DuckDB analytical spine (v_orders_clean, v_order_lines, v_customer_locations, v_seller_locations) in 0.05 seconds with 100% coordinate resolution across 99,441 customers and 3,095 sellers.
2. Spatial Hub Derivation & Detrended Innovation Covariance: Geodesic Haversine clustering synthesizes $K=5$ fulfillment corridors across Brazil, computing the exact distance tensor $D_{\text{geo}}\in {\mathbb{R}}^{5\times 14994}$ in 0.156 seconds. Cross-category demand co-movement is modeled on detrended stationary residuals via spectral PCA decomposition and a Kronecker-separable Dynamic Spatio-Temporal Model (DSTM).
3. Lohr Finite Survey Sampling & Epistemic Censored Demand Probing: To correct for São Paulo’s 47% platform volume skew during national expansion stress tests, Aegis applies Horvitz-Thompson inverse-probability weighting (${\hat{t}}_{HT}={\sum }_{i\in \mathcal{S}}{y}_i/{\pi }_i$) anchored to IBGE demographic frames. When regional stockouts right-censor observed sales ($R_{t,k,i}=0$), Bayesian Optimal Experimental Design (BOED) and Sequential Monte Carlo (SMC) particle filters actively probe latent demand intensities.
4. Fault-Tolerant Distributed Sagas & Bit-for-Bit Replay Determinism: State transitions execute through orthogonal Harel statecharts and Garcia-Molina Compensating Sagas (VerifyAuthHold -> ReserveStock -> BookCarrier -> DispatchAndSettleASC606), guaranteeing zero zombie inventory locks without distributed 2PC bottlenecks.

The 4-Policy Meta-Class Tournament

Every historical and synthetic shock trajectory is benchmarked across Powell’s four universal policy classes under identical counter-based PRNG seeds:

• Class 1 — PFA (Greedy Geodesic Nearest-Neighbor + $(s,S)$ Replenishment): Baseline operational heuristic.
• Class 2 — CFA (Parameterized Buffer MILP Tuned via Knowledge Gradient): Maximizes immediate ASC 606 contribution plus concave safety buffer valuation ${\sum }_{k,i}{\theta }_{k,i}\ln (R_{t,k,i}-x_{t,k,i}^{\text{route}}{q}_i+1)$.
• Class 3 — VFA (Post-Decision Bellman State $S_t^x$ + CAVE Dual Shadow Prices): Eliminates the expectation operator inside the $\arg \max$ by maintaining piecewise-linear concave marginal inventory value slopes ${\bar{V}}_{t,k,i}(R_t^x)$.
• Class 4 — DLA (Rolling-Horizon Lookahead + Spatio-Temporal Graph Convolutional Rollout): Predicts multi-step network congestion and transit delay propagation over the inter-hub adjacency matrix $A\in {\mathbb{R}}^{K\times K}$.