Downstream

Every change, traced to every purchase order.

Comment

Downstream

Every change, traced to every purchase order.

Inspiration

Construction projects rarely fail because information is missing—they fail because critical information lives in disconnected systems.

During our research, we discovered that an approved design change is usually well documented. It exists in RFIs, drawings, specifications, and project management software. Meanwhile, procurement teams work in completely different systems where purchase orders, vendors, deliveries, and commercial decisions are managed.

The problem isn't that information doesn't exist. The problem is that no system automatically connects these two worlds.

As a result, procurement teams often discover the impact of an approved design change only after materials have already been fabricated, shipped, or purchased, creating avoidable delays, rework, and additional cost.

Existing platforms are excellent at managing documents or supply chains individually, but there is still a missing layer that automatically translates an approved change into clear procurement actions.

That missing layer inspired Downstream.

What it does

Downstream monitors approved design changes—such as RFIs and drawing revisions—and automatically traces their downstream commercial impact.

Instead of forcing procurement teams to manually search across drawings, specifications, purchase orders, vendor records, and delivery schedules, Downstream generates a single Procurement Impact Brief that identifies:

  • Affected purchase orders
  • Impacted vendors
  • Deliveries at risk
  • Supporting evidence
  • Recommended actions

Every recommendation is backed by project documents, allowing procurement teams to review the evidence before making any decision.

Downstream never replaces human judgment—it accelerates it.

How we built it

Rather than relying on a single LLM prompt, Downstream is designed as an explainable AI workflow.

Document Intelligence

Construction documents such as RFIs, drawings, specifications, and procurement records are processed to extract structured project information.

Entity Extraction

Key project entities—including specification sections, cost codes, purchase orders, vendors, drawing references, and delivery milestones—are identified and standardized.

Knowledge Graph

These entities are connected into a project knowledge graph that represents how design decisions relate to procurement activities.

Instead of guessing relationships, Downstream follows existing project references already present across construction documents.

Construction-Specific RAG

Whenever explanations are generated, retrieval is limited to the project's own documents, ensuring every recommendation can be traced back to verifiable evidence.

Multi-Agent Workflow

Specialized agents perform document analysis, graph traversal, impact identification, and report generation before producing a Procurement Impact Brief for review.

Human Approval

Downstream only recommends actions.

Procurement teams remain fully in control of every commercial decision.

Challenges we ran into

The biggest challenge wasn't building an AI model—it was understanding how construction projects actually work.

Mapping approved design changes to procurement required understanding how RFIs, specification sections, cost codes, purchase orders, vendors, and deliveries are already connected across different systems.

Another challenge was designing a system that remains trustworthy.

Rather than allowing the model to "guess" procurement impact, we deliberately combined deterministic graph traversal with language models only where natural language understanding was required.

This made explainability a core design principle rather than an afterthought.

Accomplishments we're proud of

  • Designing an explainable AI workflow instead of another construction chatbot.
  • Building a procurement-first solution around a real industry workflow.
  • Combining document intelligence, knowledge graphs, RAG, and agentic AI into one coherent system.
  • Keeping every recommendation evidence-backed and human-approved.
  • Creating a workflow that complements existing construction platforms instead of replacing them.

What we learned

This project taught us that enterprise AI is rarely about building a smarter model.

It's about connecting the right information, making reasoning transparent, and designing systems that professionals can trust.

We also learned that knowledge graphs and deterministic workflows play a critical role in reducing hallucinations for high-stakes enterprise decisions.

What's next for Downstream

Our current prototype focuses on one high-impact trigger: approved design changes.

Future versions will expand to include:

  • Supplier delay monitoring
  • Schedule impact analysis
  • Forecasting procurement risks
  • Cross-project learning
  • ERP and construction platform integrations

Our vision is not to replace existing construction software.

Our vision is to become the missing decision layer between project changes and procurement execution.

Built With

Share this project:

Updates

Submission history