Built from the deals it was meant to solve

Minervian AI Inc. was incorporated in December 2025, and the product went live in May 2026. It is AI-powered underwriting software for commercial real estate, and it exists because of specific problems met while underwriting institutional deals — at a global developer, on large-scale urban development, and inside private equity firms managing institutional LP capital and JV waterfall structures. The modeling conventions in the product are not an interpretation of how the industry works. They are the conventions used on live deals.

Why Minervian AI exists

The last underwriting built entirely by hand before this product existed was a multi-property industrial portfolio. The broker supplied Argus exports, but the firm did not run Argus, so every property came into Excel and every property needed manual work. One tab per property. Partial sales and partial loan discharges were awkward to model and easy to get wrong.

The arithmetic of a single assumption change tells the story. Changing one inflation rate across ten properties meant opening ten files, editing each one, saving each one, re-exporting each to Excel, pasting each into the right tab of the master model, and then repairing the rows that no longer lined up.Conservatively twenty-five minutes, for one assumption. In Minervian AI that same change is a single instruction to the AI assistant and takes about fifteen seconds.

Portfolios of that shape are not unusual. Retail and industrial deals routinely carry dozens of properties, and even office portfolios commonly run to one to three buildings. And had that industrial portfolio instead been an office deal with base-year recoveries, the spreadsheet approach would not merely have been slow. It would have been impossible.

The two sections that get rebuilt on every deal

Financing and the waterfall. Every deal carries a different capital structure and a different promote, so neither section survives being templated.

Capital stacks in practice run well past a single mortgage: first mortgage, preferred equity, EB-5, foreign bonds, and bespoke bridge facilities in which each use of proceeds — capital expenditure, leasing costs, operating shortfalls — carries its own sizing, funding mechanism, and draw rate. Five tranches on one deal is not exotic.

Waterfalls are harder still, because they are negotiated rather than specified. Tiered promotes with IRR hurdles, preferred returns with kickers, catch-ups, and term sheets running to eight tiers. A template models one family of structure. It does not survive the deal changing shape.

Why the calculation engine is deterministic

This was settled by trying the other way first. Large language models get roughly 80 percent of the way to an underwriting quickly, and then stop. The remaining 20 percent is where a deal is won or lost, and the failures there are structural rather than incidental:

• Numbers diverge between runs instead of converging.
• Hallucination rates are high on exactly the line items that matter most.
• Each run takes 10 to 15 minutes before returning for more inputs and corrections.
• Nothing is surgically editable. Moving a cap rate 25 basis points means re-running everything.
• Token consumption runs to roughly 100,000 per run.
• A full monthly cash flow, which interest calculations require, stays out of reach.

There is also a harder limit. A spreadsheet produced by a language model cannot carry commercial leases with real recovery structures, renewal probabilities, and leasing commissions. That is not a prompting problem, and no amount of prompt engineering resolves it.

So the division of labour is fixed. The AI reads documents, proposes the assumptions those documents omit, edits inputs from plain-language instructions, and explains how any section calculates. A prebuilt deterministic engine performs every calculation. Identical inputs return identical numbers, every time.

The engine was validated the slow way. Dozens of test deals were underwritten twice — once in the product, once by hand in Excel — and the outputs compared line by line until they tied. That is also why the intermediate schedules are exposed as outputs rather than hidden inside a result: they are the same schedules used to check the engine against a spreadsheet in the first place.

Why the waterfall has no fixed template

Adding another tier to an existing spreadsheet template is trivial. What is hard is the back-and-forth of a live negotiation: swapping a hurdle structure for a preferred return, inserting tiers in the middle for a special bonus or promote, or replacing an IRR hurdle with a catch-up expressed as a share of deal profit. Most firms' templates assume a GP and an LP and nothing else. Deals with several LPs and co-investors need a bespoke model, and in practice only the strongest analyst on the desk can build one. A two-dimensional spreadsheet cannot be made dynamic in that way.

So when a managing director says "change it to a 15 percent preferred to the LP, GP catches up, then 18 percent to the LP, then the GP's special bonus, then 25 percent, then all residuals to the GP," that sentence can be pasted straight into the chat and the structure exists seconds later. The traditional path is two to three hours from an excellent associate, and considerably longer if the template will not bend.

Why New York conventions are built in

Because approximating them misstates the cash flow you are actually being asked about.

Leasing commissions are the clearest case. Conventional tools model a commission as a percentage of lease value. The New York convention is tiered — one percentage for the first five years, a different one for the next five, and so on. Rent escalations diverge the same way. Most markets escalate annually — a percentage or a fixed dollar amount every year. New York leases commonly step by a larger fixed amount every few years instead. The figures vary deal by deal; it is the shape of the escalation that conventional tools get wrong.

A proxy annual increase can approximate long-run cash flow well enough. What it does not reproduce is near-term cash flow — and near-term cash flow is what determines debt service coverage, the sizing of a bridge loan, and ultimately a bid.

Where the name comes from

Minerva — the Roman goddess Ovid called the goddess of a thousand works. She stands for wisdom, intellect, and reason; for strategy; for mathematics and numbers; and for the arts and crafts, trade and commerce, and skillful practical knowledge.

That combination is unusually apt for the built environment, which is at once a numbers discipline, a negotiating discipline, and a craft.

What Minervian AI deliberately does not build

Judgment about scope matters as much as capability. Several things are deliberately out of scope:

• Anything that amounts to a thin wrapper over a general-purpose model.
• Market research, zoning research, and rental comparables — a general assistant already does these well.
• Simple multifamily deals that an in-house spreadsheet handles perfectly well.
• Fast, simplified models that are not investment-memo ready.
• Lease extraction whose output is a dashboard rather than a model.

One modeling choice belongs on that list too. Multifamily value-add is projected by cohort, not unit by unit. A per-unit projection looks more precise, but it is really a claim to know which apartment turns in which month. Business plans are never written that way — they are written as an intention to turn units sitting 20 to 40 percent below market, at some pace per month — so the model takes the same shape as the plan it is modeling.

Two things are often assumed to be on that list and are not ruled out: firm-level memory across deals, and agents that keep working while you are away. Neither is offered today.

What the platform is built for is accuracy — IRRs correct to the decimal place, waterfalls that reflect the term sheet as signed, and deals where a 5 percent difference in valuation is the difference between losing the bid and overpaying.

Company details