Causal Delay Copilot

Inspiration

Construction megaprojects have a consistent overrun problem. Research from McKinsey shows that most large capital projects run significantly over budget and behind schedule, and this pattern repeats across the industry regardless of project size. The reason is not that teams lack tools to spot risk. Predictive platforms already flag delays before they happen.

The gap we noticed is what happens after the flag. A project manager sees that a shipment is at risk, but the tool does not tell them why. This is a problem we had worked on before in a different context. At Inter IIT Tech Meet 14.0, we built a causal inference pipeline using Double Machine Learning and Generalized Random Forests for Observe.AI, where the goal was the same in principle, separating a true cause from something that just moves together with the outcome. When we looked at the Supply Chain track, we researched the existing platforms in this space, including Kaya's own, and found they are built on prediction and pattern matching. None of them answer the causal question. That gap is what we decided to propose a solution for.

Our Approach

The core of the proposed system is a causal engine built with EconML, using Double Machine Learning and Generalized Random Forests to estimate the true effect of a specific factor, such as relying on a supplier who is also serving other active projects, on delivery delay. This is different from a predictive model because it explicitly controls for confounders like material type, order size, and season, so the effect it reports is not just a correlation that happens to look strong in the data.

On top of the causal engine, we plan to build a LangGraph agent layer. When a risk flag comes in, the agent would retrieve the causal driver behind it from the engine, translate that into plain language a project manager can act on, and draft one concrete action, such as an expedite request or an alternate vendor suggestion, tied directly to that driver. The causal engine finds the reason. The agent turns the reason into a next step.

We are grounding this in existing research rather than proposing something unproven. Wyrembek, Baryannis, and Brintrup (2024) showed that causal machine learning can isolate specific, actionable drivers of delay in supply chain settings, including cases where a shared supplier was linked to a measurable increase in delay once other factors were controlled for. We are adapting the same class of technique to construction procurement and scheduling data.

Challenges We Anticipate

Scoping this into a build that fits a hackathon timeline raises a few specific problems we have already thought through.

Construction procurement data with genuine confounding structure is not easy to find publicly, so we will need to construct a dataset carefully enough that it actually tests the causal method rather than just decorating a predictive one.
DML and GRF are heavier to run than a standard predictive model, so we will need to think about where the agent layer waits on the causal engine and where it can move independently, to keep the demo responsive.
Narrowing the causal question to one clean, demonstrable driver, the shared supplier effect, is deliberate. It would be easy to try to explain every possible delay driver, but we think a single well grounded result will be more convincing than a broad, shallow one.
Explaining the difference between a causal driver and a correlated one to a non-technical judge, inside a 10 slide deck and a 2 minute video, means we will have to cut a lot of technical depth and keep only the parts that make the distinction obvious.

What We Learned While Researching This

The biggest lesson so far has been in translation, both technical and communicative. Taking a causal inference approach we had built for a different industry at Inter IIT and adapting it to construction logistics means rethinking which confounders actually matter in this domain, since shared vendors, seasonal exposure, and material type behave differently in construction than in the original use case.

We also learned, through directly checking Kaya's platform and competitors, that the correlational versus causal gap is real and not something we are assuming. That gave us confidence the differentiation angle is genuine rather than a framing exercise. And in preparing this proposal, we learned how much discipline it takes to state a causal claim narrowly enough to be credible, rather than broadly enough to sound impressive.