Taking hours of manual assembly out of every appraisal report

A commercial appraisal is a legal instrument. It carries a valuation that lenders and courts rely on, and it is assembled largely by hand: comparable-property data gathered and re-keyed, narrative sections rewritten from the last engagement, tables formatted one row at a time. Appraisers were losing hours per report to work that was both repetitive and, precisely because it was repetitive, easy to get wrong.

The obvious product was a purpose-built application that owned the whole workflow. That was the wrong product. Appraisal teams already lived in Google Workspace. Their templates, their internal review cycles, their client deliverables all ran through Docs, Sheets, and Drive. Anything that required them to leave it would have been evaluated once and abandoned.

The hard constraint was set before the first line of code: automate the report without relocating the people who write it.

The architecture below was settled before build started. That sequencing is why working software reached the client within the first weeks of the engagement, with the full MVP delivered at six months: the early sprints were spent building, not renegotiating structure.

StartDeck is the hub of a hub-and-spoke system. It owns everything durable, including projects, templates, teams, identity, billing, and the appraisal database, and exposes a deliberately separate API surface for the two Workspace add-ons, isolated from the API the web application uses.

Comparable-property data, one warehouse per tenant. Commercial real estate datasets are large, irregular, and queried across geography, property type, date range, and unit economics simultaneously. That is a workload that suits a warehouse, not a transactional database. Each customer gets their own BigQuery dataset, and dataset ownership is modelled polymorphically because it genuinely varies: an account can belong to an individual appraiser or to a firm, and billing and access follow that distinction.

Identity, deliberately segmented. End users, the platform's own access to company-owned Drive assets holding shared templates, and the Workspace add-ons are authenticated separately rather than sharing one set of credentials. A customer's Drive permissions and the platform's own never sit in the same trust boundary.

eSignature brought in-house. Third-party signature services charged per envelope and owned the moment a report becomes final. We built the signing engine into the platform instead: the document is rendered to per-page HTML, signature fields are merged into it server-side, and the result is reassembled into a signed PDF.

One piece of that pipeline needed an older operating system toolchain than the main application could run. Rather than compromise the main image, it lives in its own small service container. An unglamorous decision that kept a hard dependency from dictating the rest of the stack.

AI where it earns its place. An AI layer sits over the knowledge base, giving appraisers answers drawn from StartDeck's own material rather than generic prose. Assistants are configurable by administrators rather than hardcoded, so the behaviour can be tuned without a deploy.