A deal pipeline that scores, underwrites and remembers

An investment desk sees far more opportunities than it can take: venture rounds, private-equity co-investments, credit, fund commitments, and multifamily, industrial and commercial properties, each arriving as a deck, an email thread or a phone call. The hard part is not finding deals; it is deciding them consistently, remembering why, and never letting one stall without anyone noticing. Before this application the desk tracked that in a spreadsheet and a shared inbox, and a deal's history lived in whoever had handled it.

The Deal Evaluator gives every opportunity the same path. It arrives by email or by hand and is drafted into a record; it is classified into one of seven deal types, each with its own diligence checklist, scoring rubric and weights; it moves through the stages of a pipeline where one person is responsible at every step; it is scored on questions with red-flag knockouts, underwritten against the sponsor's own numbers where the type calls for it, compared against its peers, and written up in a memo that grows with the stage. Follow-ups, sign-offs and the emails and documents that came with the deal all hang off the same record, so the page for a deal is its complete dossier.

The desk reads the pipeline board every morning, the analysts work inside the deal pages, and the principal reads the memos and the comparisons before a decision. A few rules keep it honest: a sign-off records whose court a deal is in but blocks nobody, so nothing can silently stall behind a missing signature; a memo section written at screening is carried forward rather than rewritten, so the reasoning at each stage stays on the record; and the AI only ever reads a deal's own documents, never the open web, so what it drafts can be checked against the file.

The Deal Evaluator is where inbound opportunities get decided. A deal arrives by email or by hand, is classified into one of seven types (venture, private equity, credit, a fund, or a multifamily, industrial or commercial property), and walks a pipeline of stages with a named owner at every step. Each type has its own rubric: weighted questions, hard red-flag knockouts and a computed recommendation, so two people scoring the same deal are measured against the same model.

Some of the parts worth a look:

• The underwriting module hangs a standardised apartment underwrite off a deal: every income and expense line is driven by a method (the house standard, the T-12, the sponsor's budget or a custom figure), reconciled against both source documents, and rolled into a five-year pro-forma with a refinance, an exit and the waterfall.
• The stage memo is one document that grows with the deal. A section written at screening is carried into the first assessment and the underwriting memo rather than rewritten, and the combined dossier points back instead of repeating.
• Sign-offs and ownership never gate anything, by design: anyone can move any deal, and the system records who did and whose court it is in, so nothing silently stalls without a name on it.
• The intake reads a forwarded email or a deck and drafts the record; the diligence checklist is per type, and the AI reads the deal's own documents when asked, never the open web.

Python and Flask on PostgreSQL, server-rendered pages with Alpine.js for the interactive parts. The per-type configuration — checklists, rubric questions, section and question weights, stage owners, underwriting bands — is data edited in the app, not code, so the desk can change how a type is judged without a deploy.

• Email intake reads a forwarded thread and its attachments through the Microsoft Graph API and drafts the record; the diligence checklist and the rubric are instantiated for the deal's type the moment it is classified.
• Document reading, not web search. When the AI is asked to summarise, score or draft, it is given the deal's own documents and nothing else, so every claim it makes can be checked against the file it came from.
• Scoring is explainable. A recommendation is computed from the answers through the stored weights and knockouts, and the evaluation page shows every step of that arithmetic.
• Shared platform. Single sign-on with the office's other applications, the same database cluster, nightly backups to two object stores and a deploy on every push through a serialised safe-deploy.
• Over twelve hundred tests, checked by mutation so that a test which cannot fail is found before it is trusted.

The code itself is private. If you would like to walk through it, get in touch.