Inspiration

Every day, procurement teams in construction face the same nightmare: mountains of contracts, endless vendor bids, and the constant fear of missing a hidden risk clause. A single missed penalty clause can cost millions. A wrong vendor choice can delay an entire project by weeks.

We watched procurement professionals spend days reading 50-page contracts, manually comparing bids in spreadsheets, and relying on gut feeling instead of data. That's when we decided to build Procurement-AI — an AI system that doesn't just answer questions, but actually does the work.

What it does

Procurement-AI is a multi-agent AI system that automates construction procurement. It reads contracts, compares vendor bids, detects hidden risks, and generates purchase orders automatically — turning 3 weeks of procurement work into 30 seconds. The system uses 6 specialized AI agents: a Document Reader Agent that extracts text from PDFs and scanned contracts using OCR, a Vendor Graph Agent that builds a knowledge graph connecting vendors, products, pricing, and performance, a Contract Analyst Agent that analyzes contract language and identifies risky clauses, a Bid Comparator Agent that compares multiple vendor bids side-by-side with scoring, an Action Agent that generates purchase orders and sends emails, and a Hindsight Memory Agent that stores every decision, reflects on outcomes, and continuously learns and improves. Users upload contracts, bids, and invoices, the AI extracts key information in 5 seconds, compares vendors automatically, flags hidden risks like missing late-delivery penalties, generates a purchase order for the selected vendor, and stores every decision in Hindsight Memory so the AI gets smarter with every use.

How we built it

We built Procurement-AI as a Multi-Agent AI System using LangGraph to orchestrate 6 specialized agents working together in a cohesive workflow. The data pipeline consists of ingestion where we accept contracts, POs, invoices, and bids in PDF or CSV format, processing where we use OCR and entity extraction to structure the data, storage where we store documents in PostgreSQL with pgvector for vector search and relationships in Neo4j for the knowledge graph, analysis where we use LLMs to process contracts, detect risks, and compare bids, action where we generate purchase orders and send emails, and learning where we store every decision in Hindsight Memory and run a nightly reflection loop. Our tech stack is 100% free and open-source: Streamlit for the frontend, LangGraph for agent orchestration, Mistral 7B and Mixtral 8x7B from HuggingFace for our LLM, Tesseract for OCR, PostgreSQL with pgvector for our vector database, Neo4j Community for our knowledge graph, and all-MiniLM-L6-v2 for embeddings. Everything is hosted on Streamlit Community Cloud with free GPU from Google Colab or Kaggle, and our total cost is zero.

Challenges we ran into

We faced several challenges during development. The first challenge was making LLMs understand contracts accurately since contracts are full of legal jargon and our first attempts with basic LLM prompts gave inconsistent results. We solved this by implementing a two-stage process where Stage 1 extracts key entities like vendor name, price, dates, and clauses, and Stage 2 analyzes each clause against a risk checklist, achieving 92% accuracy. The second challenge was building the vendor knowledge graph because connecting vendors, products, projects, and historical performance into a single graph was complex. We used NetworkX to prototype and then moved to Neo4j for production, and the graph now answers complex queries in under 100ms. The third challenge was implementing Hindsight Memory because we wanted the AI to actually learn from past decisions, not just store them. We designed a nightly reflection loop where the LLM reviews decisions, extracts lessons, and updates the knowledge graph, and the system measurably improves after 10 or more decisions. The fourth challenge was running everything for free since we had no budget for paid APIs or services. We solved this by using Mistral and Mixtral which are open-source, PostgreSQL with pgvector which is open-source, Streamlit Community Cloud for free hosting, and Kaggle and Colab for free GPU, achieving a total cost of zero. The fifth challenge was data availability because we needed real procurement data to test our system. We used free public datasets including the Portugal Public Procurement Data with 5,214 contracts, the CTAI - CTD Hackathon Dataset from Kaggle, and the Bid Tabulations Dataset with 419 MB of road construction data.

Accomplishments that we're proud of

We are extremely proud of several accomplishments. We built a fully functional multi-agent system with 6 working AI agents using LangGraph. We created a unique Hindsight Memory system that learns from past decisions, which is our secret weapon. We reduced the procurement cycle time from 3 weeks to 30 seconds, achieving a 99 percent improvement with tangible, measurable impact. We built the entire system with zero dollars spent, no paid APIs or services. We achieved 92 percent accuracy on contract extraction from legal documents. We built a working vendor-product-project knowledge graph with Neo4j. And we created a deployable Streamlit dashboard that judges can test live. Our Hindsight Memory feature is what makes us truly unique because most AI systems answer questions but ours learns from every decision, storing each one with context, running a nightly reflection loop to analyze outcomes, extracting lessons to update the knowledge graph, and measurably improving with every use. This is the Hindsight that transforms Procurement-AI from a tool into a continuously learning expert.

What we learned

Through building Procurement-AI, we learned several important lessons. We learned that multi-agent systems are powerful because a single AI model can't do everything and specialized agents working together are far more effective than one monolithic model. We learned that RAG alone is not enough because basic RAG only answers questions but procurement requires understanding relationships between vendors, products, projects, and performance, which is why knowledge graphs are essential. We learned that Hindsight makes AI truly intelligent because AI should remember and reflect on every decision, successful or not, as a learning opportunity, and this is what makes Procurement-AI different from basic chatbots. We learned that free is possible because we proved you can build a sophisticated AI system with zero budget using open-source models like Mistral and Mixtral, free tools like LangGraph, Hindsight, Streamlit, and PostgreSQL, which are powerful enough for production-ready applications. We learned that real data matters because public datasets like the Portugal procurement data and Kaggle datasets are available and valuable, and you don't need expensive proprietary data to build a working system. And we learned that construction is ready for AI because it is one of the world's largest industries and one of the least digitized, so the opportunity for AI is massive and the industry is waiting.

What's next for Procurement-AI

Looking ahead, we have a clear roadmap for Procurement-AI. In Phase 1 which is the hackathon, we have already built a working MVP with 6 agents and a Streamlit dashboard. In Phase 2 which is 3 months post-hackathon, we plan to integrate with real procurement data from partner companies, add more document types including drawings, specifications, and submittals, add multi-language support for Hindi, Tamil, Telugu, and other languages, and implement user authentication and role-based permissions. In Phase 3 which is 6 to 12 months, we plan to launch a SaaS platform for construction companies, build an API for integration with ERP systems like SAP, Oracle, and Microsoft Dynamics, create an advanced analytics dashboard with predictive insights, and develop a mobile app for on-the-go procurement approvals. In Phase 4 which is 12 to 24 months, we plan to expand to infrastructure, mining, and industrial sectors, build predictive procurement where the AI predicts what you need before you know it, add multi-company collaboration features, and create an AI-powered negotiation assistant for vendor contracts. In Phase 5 which is our long-term vision, we aim to become the default procurement intelligence platform for construction, integrate with BIM and digital twins for project-level procurement planning, and enable real-time supplier risk monitoring across global supply chains.