Nudge AI

Inspiration

A single delayed cement or steel shipment can stall an entire construction site for days — yet no one sees it coming until it's too late. Existing supply chain software only records delays after they happen. We noticed the real problem isn't a lack of data — it's a lack of incentive: the party that suffers the delay (the Project Manager) doesn't control the data, and the party that controls the data (the vendor) doesn't feel the pain. That misalignment is why this problem has stayed unsolved for decades, even with plenty of existing logistics tools.

What it does

Nudge AI will be an agent that predicts shipment delay risk in construction material supply chains — cement, steel, rebar — before the delay happens, and proactively alerts the Project Manager with a specific, actionable fix.

Ingest — Pulls in vendor reliability data, shipment tracking, weather, material type (standard vs. custom), order size, advance payment status, and GST filing regularity as signals
Predict — A gradient boosting risk model will score each active shipment's delay probability
Act — When risk crosses a threshold, an LLM agent (via RAG) will generate a plain-language alert with a suggested fix — reassign vendor, shift schedule — which the PM reviews and approves before execution

It's designed to be an active agent, not a passive dashboard — the differentiator is the prescriptive action, not just the alert.

We've also designed the system to give vendors a reliability score they can use to build reputation and win more contracts — turning monitoring into a shared incentive rather than one-sided surveillance.

How we plan to build it

Backend: FastAPI
Data layer: MongoDB (with vector search for RAG retrieval)
AI layer: Gemini for the agent/reasoning layer, gradient boosting (XGBoost/LightGBM) for the risk-scoring model
Frontend: PM-facing dashboard showing shipment risk overview, risk breakdown, and one-click resolve flow

Challenges we anticipate

Balancing model interpretability with predictive power — we're choosing gradient boosting over more complex temporal models specifically because judges and PMs alike need to trust why a shipment is flagged, not just accept a black-box score
Designing an incentive structure that gets vendors to actually participate, rather than resist a monitoring system
Working with synthetic data for the prototype, since real vendor/shipment data isn't available at this stage — we're building our feature set (vendor history, weather, material type, order size, advance payment, GST filing regularity) to closely mirror real-world signals

Accomplishments that we're proud of

We chose honesty over hype in our proposal. We caught and corrected several instincts to overclaim — replacing a fabricated "industry partnership" claim with real research framing, correcting inflated market stats to only numbers we verified, and choosing gradient boosting over flashier temporal models because it's provable within our timeline, not because it looks better on a slide.
We designed for adoption, not just accuracy. Most delay-prediction concepts stop at "predict the risk." We went further and designed a reliability-score mechanism that gives vendors a real incentive to participate and share accurate data.
We scoped a genuinely buildable MVP. Rather than promising to solve every material category or build a fully autonomous system, we deliberately scoped to the highest-impact materials (cement, steel, rebar) and kept a human-in-the-loop design.
We designed a feature set with real domain reasoning, going beyond the obvious signals (vendor history, weather) to include material type, order size, advance payment status, and GST filing regularity as a proxy for vendor operational discipline.

What we hope to learn

Whether honest, defensible design decisions hold up better with judges than impressive-sounding but unverified ones. We're deliberately choosing the simpler, provable model (gradient boosting) over flashier alternatives because reliability under a real timeline matters more to us than theoretical performance.

What's next for Nudge AI

Build and validate the risk-scoring model on synthetic data (Stage 2, by July 31)
Real vendor API integrations
A vendor-facing portal so reliability scores are actually visible to vendors, not just internal
Piloting with real shipment data from a construction partner