Causal decision platform

Prescriptive logistics,
backed by causal evidence.

An enterprise-grade decision engine that combines causal inference, risk-aware stochastic optimization, and a digital twin to recommend the next operational action — and certify the uplift offline before deployment.

Decision engine· live
t = 01
Causal pipeline
Observe
Estimate
Prescribe
Recommended action
Reallocate 200u
Warehouse A → B
+89%
service level
+89%
service level uplift
A/B digital twin
−81%
total cost
same scenario
95k
Olist orders
real training data
CVaR₀.₉₅
tail-risk optimal
Rockafellar–Uryasev
The problem

Logistics AI is stuck predicting the past.

Every operations team drowns in dashboards that scream “this shipment will be late” — but none of them tell the manager what to actually do, or what happens if they don't.

The right question is causal: “If I move 200 units from warehouse A to B today, what is the worst-case cost I should expect next month?” Answering that requires do-calculus, tail-risk optimization, and a simulator — not another gradient-boosted ETA.

Smart Logistics is the first open platform to wire all three together into a single closed loop with a fully auditable trail.

"How late will this shipment be?"predictive
"What is the expected delivery time?"forecast
"Cluster customers by behavior."descriptive
"What should I do now, and what if I don’t?"prescriptive
"Is the proposed decision Pareto-improving?"causal + simulation
"How will tail risk shift under policy X?"CVaR-aware
The novelty

Four novel claims, all implemented & tested.

Claim 01

Causal-informed CVaR

A DoWhy average-treatment-effect estimate modulates the stockout penalty of a Rockafellar–Uryasev CVaR linear program. First open implementation.

Claim 02

Closed-loop β calibration

Post-deployment realized costs are logged and refit via OLS, so penalty sensitivity β adapts to your operation — no manual tuning.

Claim 03

Digital-twin A/B validation

Every prescriptive decision is certified Pareto-improving in a SimPy twin before it ever reaches production.

Claim 04

Risk-aware VRP scoring

OR-Tools deterministic CVRP is re-scored with Monte-Carlo CVaR of makespan under travel-time noise for route robustness.

How it works

One closed loop. Every stage MLflow-logged.

Causal
DoWhy ATE
Penalty
β · Δ·ATE
CVaR LP
Rockafellar–Uryasev
Risk VRP
OR-Tools + MC
Digital Twin
SimPy A/B
β refit
OLS on log

Every box is a tested FastAPI endpoint. Every arrow is an integration test. Every artifact — model, scenario set, decision, KPI uplift — is persisted in MLflow as experimental evidence.

The live demo runs API + UI + Postgres + Redis on Vercel and Render. The full monitoring stack (MLflow, Prometheus, Grafana) ships with the repo and runs locally via docker compose up.

Tech stack

Boring, battle-tested, open source.

Causal ML
  • DoWhy
  • EconML
  • scikit-learn
  • statsmodels
Optimization
  • CVXPY
  • OR-Tools
  • Rockafellar–Uryasev LP
  • Monte-Carlo CVaR
Simulation
  • SimPy
  • A/B offline backtest
  • Monte-Carlo KPIs
Backend
  • FastAPI
  • SQLAlchemy
  • Alembic
  • PostgreSQL
  • Redis
  • Pydantic v2
Frontend
  • Next.js 16
  • React 19
  • TypeScript 6
  • Tailwind CSS 4
  • Radix UI
  • Recharts 3
  • react-three-fiber
AI / Copilot
  • OpenAI chat API
  • LLM narrator
  • operator Q&A
  • feedback loop
Ops
  • Docker Compose
  • MLflow (local)
  • Prometheus (local)
  • Grafana (local)
  • GitHub Actions
Deploy
  • Vercel — frontend
  • Render — API + DB + Redis
  • cron-job.org — keep-alive
Who is it for

Anyone who needs to decide, not just predict.

Supply-chain & ops teams

Replace your delivery-time dashboard with a decision copilot that recommends reallocations and proves the uplift.

3PL & last-mile operators

Get risk-aware vehicle routing that is robust to traffic shocks, not just optimal on paper.

E-commerce platforms

Causal inventory reallocation across warehouses to defend service-level under demand uncertainty.

OR / causal-ML researchers

A real, reproducible end-to-end stack to benchmark your algorithm on a public dataset (Olist).

Logistics consultancies

White-label the UI, swap the dataset, deliver prescriptive recommendations with an audit trail.

Research auditors

Every component is MLflow-logged with timestamps — ready-made experimental evidence.

Where it's going

The path from open-source repo to global platform.

Shipped
done
  • Causal engine on Olist (DoWhy + MLflow)
  • CVaR inventory LP (Rockafellar–Uryasev)
  • Risk-aware VRP (OR-Tools + MC)
  • SimPy digital twin + A/B validation
  • Adaptive β calibration loop
  • Next.js 16 + Radix UI cockpit
  • LLM narrator + feedback loop
Next
next
  • Real-time streaming ingestion (Kafka)
  • Multi-echelon inventory (plants → hubs → stores)
Vision
future
  • Federated causal learning across logistics networks
  • Causal LLM agents that execute decisions autonomously
  • Managed SaaS with private data connectors