SiteSync AI-Multi-Agent Construction Supply Chain smart

AI agents that autonomously track materials, predict delays, and auto-reschedule construction projects — turning supply chain chaos into coordinated action.

Comment

Inspiration

Construction is a $13 trillion global industry — yet 98% of mega-projects go over budget, and 77% are late. The #1 culprit? Supply chain failures. Materials vanish into a black box after ordering. A delayed steel shipment silently cascades into weeks of blown schedules, idle crews at $15K/day, and millions in overruns.

We asked: What if the supply chain finally had a brain?

What it does

SiteSync AI is a multi-agent AI system that brings autonomous intelligence to construction material tracking, delay prediction, and schedule recovery.

Agent 1: Procurement Intelligence Agent

  • Reads an unstructured vendor delay email using LLM-powered extraction
  • Extracts key fields: Material, PO Number, Delay (days), Original ETA, New ETA, Root Cause
  • Passes structured delay data downstream to the Scheduling Agent

Agent 2: Schedule Impact Agent

  • Cross-references the extracted delay against the active project schedule
  • Identifies every downstream task at risk — equipment rentals, labor crews, milestones
  • Calculates idle days and idle cost-per-day for each affected task

Agent 3: Financial Recommendation Agent

  • Compares total idle cost (if no action is taken) against the one-time cost to reschedule
  • Issues a single clear recommendation — reschedule or absorb the idle cost — with dollars saved
  • Project manager reviews and approves the recommendation with one click

Smart Command Center Dashboard

  • KPI cards showing delay length, tasks at risk, idle cost exposure, and potential savings
  • Gantt-style chart overlay comparing originally planned vs. delayed schedule dates
  • Live agent activity feed showing each agent's output as the pipeline runs
  • Final recommendation banner with Approve / Dismiss actions

How we built it

  • Multi-Agent Orchestration: CrewAI runs three specialized agents in a sequential pipeline — each agent completes its task and hands structured output to the next: Procurement → Scheduling → Financial
  • LLM Layer: A configurable LLM backend (OpenAI-compatible) handles document understanding and reasoning for all three agents, currently running gpt-4o-mini for the demo
  • Frontend: Streamlit "Smart Command Center" — a single-page dashboard with KPI cards, a vendor-email inbox preview, a live agent activity feed, and a final recommendation banner
  • Schedule Visualization: Plotly renders a Gantt-style chart comparing originally planned vs. delayed dates across every affected task, with the exact date shift labeled per row
  • Reliability Layer: Built a scripted mock mode that runs a fully working demo when no live API key is configured, plus retry-with-backoff and graceful fallback if the live API hits a rate limit mid-run — so the demo doesn't break in front of judges
  • Theming & Deployment: Locked the UI to a consistent light theme via Streamlit config (so it renders correctly regardless of a judge's device settings), deployed directly from GitHub to Streamlit Community Cloud
  • Iteration: Debugged and refined the schedule chart across several passes — fixing date-axis rendering, bar visibility, and label clipping — to get a clean, readable planned-vs-delayed comparison

Challenges we ran into

  • Unstructured vendor communication: Vendor delay emails have zero standard format — extracting material, PO number, delay length, and root cause reliably meant carefully prompting the Procurement Agent rather than relying on rigid parsing rules
  • API reliability during live demos: Sequential multi-agent calls can hit OpenAI rate limits/quota mid-run. We added retry-with-backoff and a graceful fallback to scripted output so a single failed call doesn't crash the whole judging demo
  • Cascading schedule logic: Mapping one delayed shipment to every downstream task it blocks (equipment rental, labor crew, inspection milestone) required modeling task dependencies and idle-cost-per-day, not just shifting a single date
  • Chart clarity on a date axis: Building a readable planned-vs-delayed timeline in Plotly required care — date-typed axes expect bar lengths as real time durations, and getting that right (vs. flat lines with no visible width) took a few iterations
  • Consistent UI across environments: Making sure the dashboard rendered correctly regardless of a judge's device theme (light/dark mode) required explicitly forcing the app's theme rather than relying on defaults

Accomplishments that we're proud of

  • Built a working multi-agent pipeline that can ingest a messy vendor email, extract material status, link it to a live schedule, identify downstream impacts, and propose a recovery plan — all autonomously
  • Achieved sub-minute delay detection vs. the industry average of 2-3 weeks
  • Designed a UI that construction professionals (not just engineers) can actually use

What we learned

  • The construction industry is drowning in data but starving for intelligence
  • Multi-agent architectures are transformative for complex domain workflows where different expertise is needed at each stage
  • Trust and explainability are just as important as accuracy in enterprise AI

What's next for SiteSync AI

  • Live integrations with Procore, P6 Oracle Primavera, and major ERP systems
  • Supplier scoring engine that learns vendor reliability over time
  • Multi-project portfolio view for general contractors managing 20+ simultaneous sites
  • Mobile-first alerts for field superintendents
  • Pilot program with a mid-size general contractor (in discussion)

Built With

Share this project:

Updates

Submission history