Cascade: Supply Chain Risk Intelligence for Construction

Construction projects depend on hundreds of materials arriving at the right time. A single delayed order can affect several connected activities, but teams often understand the full impact only after the delay has already spread through the schedule.

We were inspired by a simple question: What if a construction supply chain system could do more than track materials? What if it could predict delay risk, understand how that delay would affect the project, estimate the financial impact, and suggest the best response?

That idea became Cascade.

What We Are Building

Cascade is a proposed risk intelligence system for construction supply chains.

Instead of showing only whether a purchase order is late, Cascade will estimate a range of possible arrival dates. It will connect each material order to the project activities that depend on it and simulate how delays could move through the construction schedule.

Key Questions Cascade Will Answer:

Which purchase order creates the highest project risk?
How likely is the project to miss a milestone?
How many days could a material delay add?
What could the financial impact be?
Should the team expedite, change vendors, or resequence work?

Note: Every external action will remain strictly under human approval.

How We Plan to Build It

We have designed Cascade as four connected layers:

1. Data Extraction & Integration

The first layer will extract information from purchase orders, vendor emails, submittal logs, and project schedules. It will convert this information into a common format and link materials to the activities they support.

2. Predictive Modeling (Survival Analysis)

The second layer will predict material arrival times using survival analysis. This is important because active orders are incomplete observations. Training only on completed orders can create overly optimistic estimates.

A simplified version of our predictive model is:

$$\log T = \beta^T x + \sigma \epsilon$$

Where:

$T$ represents the duration of a supply stage.
$x$ contains features such as vendor, material category, location, and current status.

3. Schedule Propagation (Monte Carlo Simulations)

The third layer will run Monte Carlo simulations over the project schedule. In each simulation, Cascade will sample possible arrival dates and recalculate activity start and finish times.

This will produce a probability distribution over the final project completion date rather than a single, rigid estimate.

4. Agentic Workflow

The fourth layer will contain the agent workflow, driven by three core components:

A Watcher: Detects new information and disruptions.
A Diagnostician: Explains and quantifies the risk.
A Fixer: Proposes possible corrective actions.

Each proposed action will be evaluated using the following cost function:

$$\text{Expected Total Cost} = \text{Action Cost} + \text{Expected Delay Cost}$$

The system will automatically recommend the option with the lowest expected total cost.

What We Have Done So Far

At this stage, we have completed the core concept, system architecture, technical approach, data strategy, evaluation plan, and prototype roadmap. We have also designed the complete demo flow:

Initial State: The demo begins with a project that appears perfectly healthy.
Disruption: A vendor delay email arrives. Cascade updates the expected material arrival dates and reruns the project simulation.
Risk Assessment: The system calculates the new risk profile and highlights which downstream activities are affected.
Recommendation: It suggests actions such as expediting the order, choosing another vendor, or resequencing the work.
Resolution: Once the user approves an action, the system reruns the simulation to show how the project risk decreases.

Challenges We Expect

Data Matching & Alignment: One of the biggest challenges will be linking material records to the correct schedule activities. The same item may be described differently in a purchase order, a vendor email, and the project schedule. We plan to handle this using similarity matching, language models, and human confirmation when confidence scores are low.
Data Scarcity: Access to real construction supply chain data is highly restricted as most project data is private. To address this, we plan to build a synthetic data generator containing realistic vendors, purchase orders, supply stages, disruptions, and project activities. Because we will control the ground truth, we can rigorously test whether our predictions are properly calibrated instead of only showing a visually impressive dashboard.

What We Learned

Context is Everything: Predicting whether a material will be late is only one part of the problem. The exact same delay can have almost no impact on one project and create a major financial loss in another. The real impact depends heavily on schedule dependencies, available float, crew costs, and critical path position.
Uncertainty Matters: A single expected delivery date is not enough. Crucial decisions—such as whether to expedite an order—depend entirely on uncertainty and tail risk.
Accountability: Automation in construction must remain accountable. Cascade is designed to support human decision-making, not replace it.

What Comes Next

The next phase is to build the working prototype.

Initial Scope: A synthetic project, probabilistic arrival predictions, Monte Carlo schedule propagation, a risk dashboard, and one complete agent flow (from vendor delay email to approved intervention).
Future Versions: Support for Primavera P6 imports, alternate vendor optimization, multiple project portfolios, and long-term decision-making under fixed budgets.

Most supply chain platforms help teams see what has already happened. Cascade is designed to help them anticipate what happens next and act before the problem spreads.