Inspiration

Construction supply chains routinely miss Required-On-Job (ROJ) dates even when purchase orders are placed on time. Delay signals are scattered across POs, submittal logs, vendor emails, and shipping updates with no unified, predictive view. Indian construction projects specifically suffer from poor lead-time estimation and fragmented tracking, cascading into missed schedules and budget overruns. We wanted to build something that predicts delays before they happen, not just reports them after.

What it does

ROJ Guard is an AI agent that predicts the probability each critical material line item will miss its ROJ date, then automatically initiates mitigation — vendor escalation, resequencing suggestions, or alternate supplier discovery — before the delay hits the project schedule. It flags at-risk materials weeks in advance with a clear Low/Medium/High risk score and a plain-language explanation of why.

How we built it

We designed a four-layer pipeline. Layer 1 (Data Ingestion) pulls in POs, submittals, vendor communications, and shipping data. Layer 2 (Project Graph & Feature Store) uses LLM-based document intelligence to extract entities and link them into a unified graph, while computing vendor reliability features. Layer 3 (Prediction & Risk Scoring) forecasts lead times using gradient-boosted models combined with domain rules to classify ROJ risk. Layer 4 (Agentic Action Layer) converts high-risk flags into concrete next steps — vendor escalation, resequencing, or alternate supplier search — with a human-in-the-loop approval gate before any action executes. This is designed to plug directly into Kaya's existing Amber project graph rather than duplicate it.

Challenges we ran into

Synthesizing realistic vendor and lead-time data without access to real Kaya datasets was difficult within the hackathon timeline. Balancing full autonomy with trust meant we had to design explicit human-approval checkpoints rather than letting the agent act unsupervised. Structuring unstructured documents like PDFs and vendor emails into a consistent graph schema also required careful entity-linking logic.

Accomplishments that we're proud of

We designed a four-layer architecture that plugs directly into Kaya's existing project graph instead of building a competing system. We combined document intelligence, forecasting, and agentic workflows into one coherent pipeline rather than just wrapping a chatbot around a prompt. We also grounded our design in real research on construction delay causes and lead-time forecasting methods.

What we learned

Agentic workflows need human-in-the-loop gates to be trustworthy in high-stakes procurement decisions. Predictive risk scoring is far more valuable to construction teams than reactive dashboards, since it gives them time to act. Combining probabilistic model outputs with domain-specific rules builds more trust with users than a black-box model alone.

What's next for ROJ Guard

Because the system operates on structured events and features rather than project-specific hardcoded logic, it can scale across any number of job sites or vendors by simply feeding in new data streams. Next steps include integrating with real Kaya project data, building out the full agentic escalation workflows, and piloting on a live construction project to validate delay prediction accuracy.