RetroRate VHS — surfacing hidden assumable mortgages while you browse
RetroRate · 2025–2026 · Owned end to end · Live on the Chrome Web Store
Assumable mortgages, FHA and VA loans a buyer can take over at the seller's
original rate, are invisible on every major listing portal, which means millions
of homes carrying a 3% loan look identical to everything else on the market. VHS
surfaces them in place: as you browse Zillow, Redfin, Realtor.com and Compass, it
flags listings with an assumable loan and shows the rate, the monthly savings, and
the lifetime interest savings against today's market, across all 50 states.
As the extension scaled to more traffic and more portals, calling out to a preview
endpoint for every listing became a bottleneck, and fragile, since a portal changing
its page could break it outright. I rebuilt it to read each portal's own
listing data directly as it loaded, which fixed both: faster, and resilient to
redesigns that used to take it down overnight. From there I took it national, solved
the throughput and race-condition problems that came with checking thousands of
listings at once without hammering our backend, and built the internal dashboard we
used to see how people were actually using it.
Chrome Web Store
Product page
TypeScript · React · Vite · Chrome Extensions (MV3) · Node
AI-assisted loan verification, from data model to certificate
RetroRate · 2026 · Concept to production
Verifying that a listing's loan is actually assumable, and at what rate, used to mean
trusting whatever the seller's agent claimed. This product turns that claim into
proof: an agent uploads a mortgage statement, Gemini and Vertex AI extract the terms
and reconcile them against a county-recorder cross-check, and a confidence-scored
certificate comes out the other end, with anomaly flags for anything that doesn't
line up. The certificate is public and shareable by QR code, and expires
automatically the moment a listing sells or goes off-market, so what's out there
always matches reality. The real challenge was trusting an AI extraction pipeline
enough to stand behind its output publicly, which meant building a rules-based
confidence engine and a full audit trail instead of taking the model's answer at
face value. Verification became a paid feature for listing agents, opening a new
revenue line for the business.
TypeScript · Node · Gemini · Vertex AI · GCP · Stripe
MLS partner integrations, from signed deal to live product
RetroRate · 2025–2026 · Concept to production
Every MLS partnership we signed was a new revenue relationship, and each came with
its own login flow, its own listing feed, and its own rules about what we were
allowed to say in an email. I built the system that turned a signed deal into a
running product: partner-specific OIDC login, a queue that turned raw MLS listings
into compliant, on-brand alert emails, and a per-partner rule catalog so each MLS
could have its own thresholds and disclosures. On top of the per-listing alerts I
set up an MLS-wide weekly digest campaign; together they now send more than 50,000
emails a month.
TypeScript · Node · SendGrid · OIDC · Cron / Queues
Turning a listing classifier into a paid agent product
RetroRate · 2025–2026 · Concept to production
Most MLS listings never mention that their loan is assumable, agents don't know to
say it, or don't know it matters. I built a model that reads MLS remarks and flags
likely assumable mentions, then used that signal to promote and rank those listings
above the rest. That became the seed of an agent-facing dashboard where agents can
see how their listings are performing, claim and verify them, and see the savings
they're marketing to buyers. Wiring agent identity to MLS IDs so the dashboard
showed the right listings, and building the org-level upsell that turned one agent's
usage into a brokerage-wide sale, were what turned a detection feature into a second
revenue line.
TypeScript · Node · React · Stripe · MongoDB
Taking search and maps from a prototype to the core product
RetroRate · 2025–2026 · Proof of concept to production
The map and search experience is the product most users touch first, and it started
as a Google Maps proof of concept. I took it to production: split list/map views,
status-colored markers, geocoded location search backed by a nationwide dataset, and
county- and ZIP-level search for buyers who didn't have a single address to start
from. The unglamorous part was keeping it working everywhere it needed to: a long
tail of responsiveness, Safari, and mobile fixes, plus the state bugs that only
surface once real users stack filters in combinations no spec accounted for.
TypeScript · React · Google Maps · Node
Replacing an SFTP file drop with a real API
Relay (formerly Kompas Care) · 2024–present · Contract
Advised a small business on replacing their manual SFTP file-ingestion process with
a REST API, then designed and built it in C#/.NET. I migrated their existing
customers onto the new interface and onboarded new ones, replacing bespoke
per-customer integration work with a repeatable process and documentation written
for non-technical stakeholders — so onboarding no longer depended on me being in
the room.
C# · .NET · REST
An offline-first field app, and the GraphQL layer underneath it
GE HealthCare · 2021–2025 · Tech lead
FX2 is a business-critical app for field teams who can't count on having a
connection, so it had to work fully offline across iOS, Android and Windows. I was
tech lead from stack exploration and proof of concept through delivery, and built its
Apollo GraphQL API layer from the ground up, serving more than 10,000 users globally.
That layer was also the company's first GraphQL integration with its major CRM
systems, so part of the job was getting teams who had only ever exposed REST to agree
on a new contract. Outside the app itself I maintained the cloud infrastructure behind
several critical app suites, migrated an AWS-hosted application to cut its monthly
cost by more than 60%, and mentored two engineers delivering features alongside me.
GraphQL (Apollo) · React Native · React · Java · AWS · Docker
An automation practice that saved $1M a year
GE HealthCare · 2019–2021 · Digital Technology Leadership Program
The leadership program rotated me through technical product management, full stack
engineering and network architecture. In the product rotation I kept hitting the same
manual processes duplicated across teams, so I pioneered an automation foundry to
attack them and built the internal site that got other teams actually using it. The
projects it delivered saved more than 20,000 hours and about $1 million annually. A
later rotation had me testing and deploying firmware updates to thousands of live
devices across active hospitals.
React · Angular · Java · SQL · Jenkins