Reel Estate

Virtual staging is an old idea with a new constraint: the moment the image is generated rather than photographed, it stops being a marketing asset and starts being a regulated disclosure. Most of the engineering here is downstream of that.

The pipeline

A job takes a listing photo, runs it through image staging, and — when the agent wants motion — hands the result to image-to-video generation. That chain is slow, expensive, and fails in interesting ways, so none of it runs inline. Work is queued, assets land in S3, and status returns by webhook. The front end never waits on a model.

Billing that survives concurrency

Credits are spent with atomic MongoDB operations behind guard clauses. The naive version of this — read balance, generate, decrement — double-spends the moment an agent fires two generations at once, which they do constantly. The guard clause is the whole feature.

Tooling the codebase itself

We run a custom MCP server for Reel Estate, so agentic tooling works against the platform’s own context rather than guessing at it from source alone. It sits alongside integrated Atlassian, YouTrack, and GA4 servers and a library of custom Claude skills that encode house style, MLS compliance rules, and Next.js scaffolding patterns — which is what keeps AI-assisted contributions idiomatic instead of merely syntactically valid.

What I own

The technical roadmap, the compliance subsystem, the acquisition stack, and production DNS across Route 53 and Vercel. I work directly with the CEO on roadmap, hiring, and enterprise brokerage deals.

The platform currently serves dozens of agents across firms including Compass, eXp Realty, and Dave Perry-Miller.