NextDealIQ

The problem

An agent working a deal moves the same figures through a CRM, an analysis spreadsheet, a report document, an inbox and a calendar. None of them knows the others exist, so the agent is the integration layer: copy the numbers out of the spreadsheet, paste them into a document, fix the formatting, export the PDF, and do it again on the next deal. The public site prices that stack at $310 to $800 a month across five subscriptions.

Investor-facing agents have it worst. The same house evaluated as a long-term rental, a BRRRR and a flip needs three separate models: the cash flows differ, the exit assumptions differ, and “return” does not mean the same thing across them. In a spreadsheet that is three files with copy-pasted assumptions, and when a rate changes some of them get updated and some don’t.

What it does

The property is the top-level record. Analyses attach to it, each typed as rental, BRRRR, flip or CMA and each carrying its own complete assumption set, so the same house can be underwritten several ways at once without one strategy’s inputs moving another’s. A property belongs to an agent and can be linked to a client, and clients move through a nine-stage pipeline from new lead to closed.

Client reports are built from the analyses rather than exported from them: the PDF payload is produced by re-running the calculations against the stored analysis, so the document and the analysis cannot disagree. Reports go out as a PDF or a share link.

An assistant sits on top. Every tool it can call has a fixed tier: reads run silently, reversible writes run and are reported, and anything destructive or visible outside the agent’s account (deleting, creating a share link, closing a client) is held as a pending action until the agent approves or rejects it. The tier lives in a server-side registry, not in the prompt, so text in a client note cannot promote a risky action into an automatic one.

How it got here

It began in 2023 as an underwriting tool for investors. In March 2026 it pivoted to the agents who serve them, keeping the calculation engine and adding the CRM, reports and assistant around it. In August 2026 the backend moved from Prisma, PostgreSQL and NextAuth to Convex and Convex Auth.

It is a 2-person team, and the product is my build: the original property and analysis app (the property and analysis creation flows, the analysis dashboards and sidebar, the rental calculator), the pivot’s business plan, roadmap and feature specs, the CRM, client reports and the AI assistant, the move to Convex, production hardening and the move to Vercel (request logging, bot tracking and API rate limiting, a diagnostics endpoint), and the public marketing site.

Where it is

The software works end to end: model a property, run its analyses, build a client report, work the pipeline. nextdealiq.com is live with a founding-member price and a waitlist, collecting interest while onboarding is finished and the numbers earn enough confidence to put in front of someone about to spend money.

That distinction carries weight here. A dashboard that is wrong by a rounding convention still looks plausible, and the person reading it is deciding whether to buy a house. Finished engineering and a ready product are two different states, and this one is in the first.

How it was built, in full

The pivot was written as documents before code: a business plan, a phased Gantt timeline, feature specifications and workflow diagrams. Each AI-assisted build phase then had a written scope and a completion record to check against. The public site was built with agent passes for the breadth (sections, charts, scroll reveals), and every claim on it was reviewed against the plan.

Integration points are written down: what the API expects from the session, what the dashboard expects from the data. Reading the code to work it out is slower and goes stale. Test coverage concentrates on the calculation layer, where a silent error is both expensive and invisible.

What stays human: every underwriting formula is checked against a worked example before it ships, because an AI-written cap-rate calculation that is off by a rounding convention still looks plausible on screen, and someone may buy a house on the strength of it. Auth and data isolation between agents are reviewed by hand rather than trusted to generated code. And launch waits on confidence that the numbers hold up in front of a client; working software and a ready product are two separate questions.

What I’d revisit

The financial models are correct, and they are hard-coded. A user who wants to underwrite with an assumption we did not anticipate has to wait for us to add it. Turning the strategy models into data instead of code is a bigger rewrite than it sounds, which is why it has not happened yet.