Empire State Realty Trust (ESRT)

Overview

At Empire State Realty Trust, the owner and operator of the Empire State Building and a broader NYC commercial and residential portfolio, I built AI tools for lease administration and natural-language lease search. I built an Excel/VBA Letter of Credit expiry notification system with deduplication logic and automated, building-routed Outlook alerts, and used Harvey AI to generate two types of lease abstracts: RAG-optimized for semantic retrieval, and verbatim for Yardi Voyager data entry. I also designed and built a working prototype of Lease Intelligence, a conversational AI tool giving non-specialist teams across Asset Management, Legal, Finance, and AR direct plain-English access to commercial lease data, layering a Claude (Anthropic API) synthesis engine over Harvey-managed retrieval on a FastAPI + React stack, deployed on Azure App Service and Static Web Apps and secured via Entra ID and Managed Identity within ESRT's Microsoft tenant.

What I did

• Built an Excel/VBA Letter of Credit expiry notification system with deduplication logic and automated Outlook alerts, deployed team-wide with building-to-email routing and user documentation.
• Built Harvey AI Agents and Playbooks generating two abstract types from lease documents: RAG abstracts optimized for semantic retrieval and verbatim abstracts structured for Yardi Voyager data entry.
• Designed and built a working prototype of Lease Intelligence, a conversational AI tool for querying ESRT's commercial lease terms in plain English.
• Layered a Claude (Anthropic API) synthesis engine over Harvey-managed retrieval, on a FastAPI + React stack deployed to Azure App Service and Static Web Apps, secured via Entra ID and Managed Identity.

Highlights

• Delivered a working Lease Intelligence prototype giving 50+ employees self-serve access to lease data previously gated behind Lease Administration, covering all 880 leases.
• Presented the production architecture (Harvey exports, Informatica ingestion, Azure storage with field-level protection, FastAPI retrieval) and rollout plan to ESRT leadership.
• Modeled an 86% licensing-cost reduction versus per-seat Harvey licenses ($227.3K vs $4.9–32.8K per year).
• Cut lease abstraction time by roughly 50% with Harvey Agents and Playbooks producing dual-format abstracts for semantic retrieval and Yardi Voyager entry.
• Deployed an automated LOC expiry notification system team-wide, cutting up to 2 hours of manual review weekly.
• Secured the Lease Intelligence platform end-to-end within ESRT's Microsoft tenant using Entra ID and Managed Identity.

What was hard

• A live production deployment was out of scope: cybersecurity review and system dependencies could not clear within one summer. I scoped the prototype to a sandbox with production-shaped architecture instead, and documented a clear path to future implementation.
• My direct manager left unexpectedly partway through the internship. I kept the work moving and the quality up without day-to-day guidance.
• The abstraction bottleneck was mundane but real: OCR'd PDFs, formatting cleanup, and field-by-field Yardi entry. Getting Harvey to produce verbatim, Yardi-shaped tables from messy source documents took careful playbook design, with human review kept in the loop.

Skills