Portfolio Intelligence

Context

Project and building managers previously sent updates through several handoffs. Operations, executive support, and analytics staff then manually combined project progress, revenue, costs, vendor activity, property updates, and operational issues whenever leadership requested a report.

Problem

Important updates could arrive late, leaving leadership with gaps in strategic visibility and less time to respond to delayed milestones, unexpected costs, revenue concerns, and projects falling behind plan.

What I built

I shipped the first working version in five days using Claude 3.5 Sonnet, Microsoft Power Automate, MCP-based integrations, Microsoft Word, Excel and Outlook APIs, structured JSON, and email-based human approval. Later iterations improved the templates and report structure while preserving the original workflow architecture.

How it works

1. 01Collect standardized updates from project and building managers whenever milestones, delays, costs, or risks change.
2. 02Retrieve supporting revenue, plan, cost, conversion, vendor, and portfolio data through MCP integrations and APIs.
3. 03Normalize the inputs into structured JSON and organize them by property, project, business area, and reporting period.
4. 04Use branching logic to compare actual progress with plans and route higher-risk issues through additional checks or follow-up requests.
5. 05Validate that every required property, project, financial signal, risk, and action item is present; retry or request missing information when needed.
6. 06Email the draft to the executive assistant for approval, incorporate requested changes, and deliver the approved weekly report to leadership.

Outcome

• Saves approximately 16 hours of manual reporting work each week.
• Supports portfolio reporting across more than eight properties.
• Replaced request-based reporting with a dependable weekly operating rhythm.
• Reduced manual handoffs and helped leadership identify bottlenecks earlier.

What I learned

A dependable reporting agent needs more than good prose: structured inputs, risk-aware branching, completeness validation, retries, and a clear human approval boundary make the output operationally useful.