Harborstone Properties

Mobile application

The challenge

Harborstone ran a $2.8B portfolio on email threads and spreadsheets. Deal documents lived in individual inboxes, valuation models were rebuilt by hand each quarter with assumptions nobody could reconstruct, and resident maintenance requests arrived by phone to a regional office that forwarded them by email. Average maintenance response was 6.4 days, and exit interviews put it among the top three reasons residents left.

Research

• Traced twelve completed acquisitions end-to-end to map the real process
• Inventoried 60+ valuation spreadsheets to identify common model structure
• Analyzed 18 months of maintenance tickets for response-time root causes
• Resident interviews across six markets on renewal decision factors

The solution

Deal flow moved into a structured pipeline with document extraction pulling key terms from LOIs, leases, and inspection reports directly into the record — the extraction is reviewed rather than trusted, but it eliminated the manual re-keying that had been the main source of both delay and error. Valuation became a versioned engine where assumptions are explicit and scenarios are reproducible. The resident app connects maintenance requests directly to dispatch, removing the regional office from the path entirely.

UX decisions

Features

• Structured deal pipeline with reviewed document extraction
• Versioned valuation engine with explicit scenario modeling
• Resident app for payments, maintenance, and community updates
• Direct maintenance dispatch with photo-first intake
• Portfolio analytics with exception-led defaults
• Investor reporting with drill-through to property detail
• Multi-jurisdiction lease term and compliance handling

Technology

Architecture

A Node service layer fronts PostgreSQL holding deal, property, and lease records, with Elasticsearch indexing documents and extracted terms for retrieval. A Python service handles document extraction and the valuation engine, where every model run is stored with its full assumption set so any past valuation can be reproduced exactly. The Next.js resident app and internal platform share the service layer with role-scoped access. The estate runs containerized on AWS with per-market data residency handling for Canadian properties.

Results

Deal cycle time from LOI to close fell 41%. Maintenance response dropped from 6.4 days to 2.1, and lease renewal improved 18% — the two moved together closely enough that the operations team now treats response time as a retention lever. Quarterly investor reporting, previously a two-week scramble, now generates in an afternoon.

Lessons learned

1. 01Document extraction succeeded because we designed it as an assistive step with mandatory review, not as automation. The fully-automated version we prototyped was rejected by analysts within a week.
2. 02Photo-first maintenance intake was a small decision with outsized effect. It improved triage more than any workflow change we made.
3. 03Reproducible valuations mattered less for analysis than for arguments. Being able to show exactly which assumption changed between two quarters ended a recurring source of internal friction.