Inspiration

We began with Track 3 and a simple question: how hard can construction procurement really be? After a few days of research, we found ourselves deep in complex issues. Long-lead electrical equipment, such as transformers and switchgear, that used to ship in under a year is now quoted at two to five years. The most surprising part was that this equipment accounts for less than 10% of a project's cost but nearly 100% of its delay risk. A completed building worth hundreds of millions can sit unused, waiting on just one component.

One of us studies materials science, so checking a product against a written specification is straightforward. Realizing that an entire industry performs this task by hand during the "submittal approval" stage—the crucial point that determines billion-dollar schedules—made the idea come together.

What it does

ProcShip serves as an AI decision layer over the procurement process. It closes one loop: Select → Approve → Track → Learn.

Select: It reads incoming vendor quotes and compares them systematically, assigning a reliability score to each vendor based on past performance.
Approve (our hero): When a vendor's submittal arrives, ProcShip automatically checks it against the project specification. It extracts every requirement (voltage rating, dimensions, standards, clearances), assesses each one as pass, fail, or needs review, cites the exact source line for each finding, and suggests approve, reject, or revise-and-resubmit—automatically drafting the follow-up for the vendor if a revision is needed.
Track: Approved orders are scheduled backward from the target energization date. If anything slips, an immediate alert is triggered.
Learn: Completed orders feed real delivery performance back into the vendor score, making the next decision sharper than the last.

The main idea is this: you can't use software to shorten a factory's lead time, but you can reduce the lead time for decisions and defend the one gate that the buyer controls.

How we plan to build it

The technology stack includes Python and FastAPI for the backend, and React for the frontend. For the submittal compliance checker, we break down the specification section and the vendor datasheet into a vector store. We use retrieval (RAG) to find the exact clause relevant to each requirement and make a structured LLM call to return typed JSON, which includes the requirement, declared value, verdict, and source line. A comparison layer creates a cited pass/fail report and a recommended decision. Every LLM call is restricted to structured output, ensuring no free-form content reaches the report and each result is traceable.

Challenges we ran into

The toughest challenge was not technical—it was resisting the temptation to create something flashy. Every instinct pushed us toward adding another dashboard or marketplace, but focusing on a single practical solution required discipline.

Accomplishments that we're proud of

We identified a clear, non-obvious problem—submittal approval—that sits on the critical path and for which almost no one is developing a solution.
We crafted a thesis that holds up under scrutiny: we are honest about what AI can’t change (the factory) and specific about what it can (the decision), which we believe is more convincing than claiming to eliminate delays.
We designed an architecture that can be applied across various equipment categories; only the specifications change.

What we learned

The biggest takeaway: in an AI-focused project, deep domain knowledge is more valuable than clever models. Our advantage came from understanding why submittals influence production timing—not from any advanced technology in the pipeline. We also realized that narrowing our focus is a strength, not a weakness. Three fully functional features tell a much stronger story than ten that only work partially.

What's next for ProcShip

Our next step is Stage 2: strengthening the compliance checker, fully developing the Select → Approve → Track → Learn loop, and adding a vision-language component so ProcShip can read graphical shop drawings as well as text datasheets.

After that, we want to expand coverage across equipment categories, base the reliability score on real historical order data, and connect to the vendor and ERP systems that procurement teams are already using. The real test will be placing it in front of an actual procurement team and seeing how much decision-making time it can save them.