ProcureGraph AI

Inspiration

Construction procurement still runs on manual document review — RFQs, BOQs, contracts, quotations, and invoices scattered across emails and PDFs. Teams are left asking the same questions on every project: What's been approved? What's delayed? Will it actually arrive on time? Nobody has good answers until it's too late, and one missed mismatch — a vendor quote that quietly promises 21 days when the contract says 14 — can silently slip an entire schedule. We wanted to build something that catches that gap before it becomes a delay, not after.

What it does

ProcureGraph AI reads construction procurement documents, links them into a live relationship model of vendors, materials, specs, purchase orders, and milestones, and automatically detects mismatches between them — an invoice that doesn't match its PO, a quote that deviates from an approved spec, a delivery promise that conflicts with the contract.

When it finds one, it doesn't just raise a flag. It generates an Evidence-to-Action Playbook:

The risk statement
The evidence chain behind it
The impacted milestone
A confidence score
2-3 ranked next actions — draft an escalation, check an alternate vendor, push an approval — one click away

A built-in RAG assistant also lets the team ask natural questions about their own project documents and get grounded, cited answers.

Instead of a chatbot bolted onto a pile of PDFs, the product is a spatial relationship canvas — vendors, items, specs, and milestones as connected nodes, with risky connections visually highlighted, so a project manager can see why something is risky, not just that it is.

How we built it

We designed the system as a pipeline: Ingest → Understand → Detect → Explain & Act → Show.

Document Intelligence: layout-aware parsing (OCR + VLM support) extracts structured fields — vendors, items, quantities, delivery terms, spec references — from raw procurement documents.
Relational Reasoning: extracted entities are linked into a lightweight dependency graph (vendor ↔ item ↔ spec ↔ PO ↔ milestone), so the system reasons across documents instead of treating each one in isolation.
Mismatch & Anomaly Detection: a rule-based comparison engine diffs linked fields against each other and scores the resulting risk.
Evidence-to-Action Playbook: flagged risks are traced through the graph and turned into a structured, explainable recommendation with ranked mitigation actions.
RAG Assistant: the same document embeddings power a grounded Q&A layer over the project's own documents.
Frontend: a canvas-first interface (built with React Flow) makes the relationship graph the main screen rather than a hidden tab, styled with a restrained, high-signal palette where color always means something — red for risk, amber for watch, teal for clear.

Challenges we ran into

The biggest challenge was scope. Procurement touches vendor scoring, forecasting, anomaly detection, agentic workflows, and reporting — it would be easy to try to build all of it and end up with nothing working well. We deliberately cut the feature list down to five that form one clean causal pipeline, and replaced full ML-based forecasting with a transparent rule-weighted risk score for this stage, with true probabilistic forecasting scoped as a clear next step rather than something we'd overclaim. Sourcing realistic procurement documents solo, without an existing construction dataset, was also a real constraint we had to design around.

Accomplishments that we're proud of

We're proud that ProcureGraph AI doesn't stop at "here's a risk" the way most document-AI tools do. By tracing every flagged mismatch through the relationship graph into a fully explainable Playbook — evidence, impact, confidence, and ranked actions — the system closes the loop between detection and decision, which is exactly the gap current construction-AI research identifies as underserved.

We're also proud of the discipline in the build itself: rather than chasing every possible feature, we scoped the product down to five features that form one clear, demoable pipeline, and designed the interface so the relationship graph — our core differentiator — is the first thing a user sees and interacts with, not a hidden tab.

What we learned

Research on AI in construction risk management consistently shows the field is strong at identifying risk but weak at acting on it — response and monitoring stages are underrepresented compared to prediction. That gap became the core justification for the Playbook feature: detection alone isn't the hard part anymore, explainable, evidence-backed action is.

What's next for ProcureGraph AI

Upgrade rule-weighted risk scoring to a proper probabilistic forecasting model once there's enough project data to train on
Expand the relationship graph from an in-memory model to a persistent, multi-project graph database
Add standalone vendor intelligence scoring as its own module
Build out the full command-center dashboard beyond the canvas and playbook drawer