Pulse AI NYC Hackathon Volume.5 - v0 by Vercel

Build an Autonomous AI Agent That Protects Construction Profitability. Overview You are the CFO of a $50M/year commercial HVAC contractor. Last quarter, three projects closed at 6.8% realized margin against a 15.2% bid margin. This is no longer an anomaly — it’s a pattern. Dashboards show data. You need something that thinks. Your mission is to build an AI agent — not a chatbot — that autonomously monitors a portfolio of HVAC construction projects, identifies margin erosion risks, investigates root causes, takes action, and reports back. The agent must: Scan the entire portfolio Detect margin risk signals Investigate by chaining tool calls Produce actionable outputs Support follow-up questions with memory Communicate findings clearly in business language What Makes This Different This is not a dashboard challenge. This is not a chatbot challenge. You are building an autonomous AI agent using: An LLM (reasoning brain) Tool calling (data access + calculations) Memory A looping mechanism (stopWhen) Email capability (proactive reporting) Your agent should continue investigating until it understands the situation — not stop after one query. Dataset Participants receive a realistic construction portfolio dataset (~18K records), including: contracts.csv sov.csv sov_budget.csv labor_logs.csv material_deliveries.csv billing_history.csv billing_line_items.csv change_orders.csv rfis.csv field_notes.csv (~1,300 unstructured reports) Embedded within the data are real-world signals: Labor overruns Scope drift Verbal approvals Billing lags Pending change orders RFI-related exposure Not every project is failing. Not every risk is obvious. Your agent must find the story. Required Capabilities Given a prompt like: "How’s my portfolio doing?" Your agent should autonomously: Assess overall margin health Identify high-risk projects Investigate root causes (labor, materials, billing, change orders, RFIs, field notes) Quantify financial exposure Recommend specific recovery actions Send an email summary or alert Support follow-up conversation with context memory.