Inspiration Construction procurement is a multi-trillion dollar industry that still runs on fragmented emails, messy PDF bids, and disconnected spreadsheets. While managing a large-scale project, teams are left asking the same questions: What's been approved? What's delayed? Are the vendors compliant with safety codes?
We realized that modern LLMs are intelligent, but they are purely "reactive"—they sit in chat windows waiting for prompts. Real enterprise procurement doesn't need a chatbot; it needs a proactive, autonomous team that can securely navigate multi-departmental debates, policy checks, and human approvals. That realization birthed FlowPilot AI.
What it does FlowPilot AI is a multi-agent society designed to act as an autonomous back-office procurement team for construction firms. Instead of relying on a single AI model, FlowPilot deploys a society of specialized agents:
Procurement Planner: Decomposes high-level requests (e.g., "Source 50 tons of rebar"). Vendor Analyst: Discovers regional vendors, compares pricing, and analyzes bids. Compliance Checker: Cross-references proposed purchases against safety codes, contracts, and strict budget limits. These agents discuss, debate, and reach a secure consensus. For high-stakes decisions (like exceeding a PO limit), the system seamlessly routes the context to a human Project Manager for a one-click approval before executing the final purchase order.
How we built it FlowPilot AI was built with a strong focus on enterprise-grade architecture:
Qwen-Max: Powers our core multi-agent reasoning, allowing the Planner and Compliance agents to debate complex policies without hallucinating. Qwen-VL: Acts as the vision engine for our Vendor Analyst agent, capable of ingesting and parsing messy, unstructured supplier invoices and contract PDFs. Frontend: We built a highly responsive, premium UI using HTML5, CSS3, and Vanilla JavaScript, animated with GSAP to visually demonstrate the multi-agent terminal workflow and real-time operations. Challenges we ran into The hardest part was ensuring Multi-Agent Consensus. It's easy to get one AI to generate text, but orchestrating multiple agents (where a Compliance Agent can overrule a Vendor Analyst) required complex state management and strict system prompting. Additionally, designing a UI that made a complex, invisible "agentic workflow" feel transparent and trustworthy to a human user was a massive design challenge.
Accomplishments that we're proud of We are incredibly proud of the Transparent Autopilot interface. By building a live terminal UI that logs every agent interaction, policy check, and decision, we solved the "black box" problem of AI. Enterprise managers can actually audit why the AI made a procurement decision.
What we learned We learned that Qwen Cloud's function calling is exceptionally reliable when properly scoped. By giving agents strict boundaries and distinct personas (e.g., strict compliance vs. cost-saving procurement), the overall system becomes significantly more resilient to errors than a single, monolithic model.
What's next for FlowPilot AI Our next step is integrating FlowPilot directly into industry-standard construction ERPs like Procore and Autodesk Construction Cloud, allowing the agents to not just draft POs, but physically execute them and track supply chain delivery statuses in real-time.
Log in or sign up for Devpost to join the conversation.