Data & research

Deal Desk

A deal command center for multifamily real estate — drop in a rent roll and a T-12, and it finds deals, underwrites them, and keeps every one of them organized from first sighting through years of ownership.

Multifamily real estate Python + LibreOffice Claude Code agents Mock mode by default

See it work

Deal Desk — live demo Open full screen ↗

The problem

A rent roll and a T-12 arrive in every format imaginable, and turning them into numbers a fund can trust takes real work every time — before anyone gets to the actual question, which is whether the deal is worth underwriting at all.

A fund looking at deals every week needs one place where each deal lives from the day it's sourced through years of ownership afterward — not scattered across email threads and one-off spreadsheets that don't talk to each other.

Running that process through AI tooling gets expensive fast if every document, every underwrite, and every research pass hits a metered API. Cost control had to be part of the design from the start, not something bolted on after the fact.

What it does

Fund desk (dashboard)
The home view — "Pipeline by stage," "Recent research runs," "Recent underwritings," and "The box," all on one screen.
Research
Where a deal's early research work lives before it moves to underwriting.
Underwriting
Runs the numbers on a deal and produces a verdict line — the prototype's example reads "Worth the full underwrite."
Pipeline
"Every deal we've touched," organized from first sighting onward.

How it's built

Two modes, set by DEALDESK_MOCK. Mock mode is the default: it produces real downloadable files with clearly-labelled placeholder numbers, runs in seconds, costs nothing, and needs no API key.

In real mode, document extraction runs on the metered API at roughly $0.15 a deal. The maths, the workbook, the memo and the one-pager are all free — plain Python and LibreOffice. The research crew runs on the Claude Code subscription, so it carries no per-token charge.

The API key itself is stripped from the process environment at startup, so a crew run has no billable credential it could reach even by accident.

Per-deal workspaces survive restarts; deleting the folder deletes the deal. Deal data is never committed to version control, because it's PII.

What we'd do next

Deal Desk is active, in-use software at Westfork Capital, with a live handoff document rather than a finished spec — new views and refinements get added as the fund's own workflow changes, not on a fixed release schedule.

See what this could look like for you

Every project on this site is real software, built the same way. Tell us what's slow or manual right now.

Start a conversation