KeystoneOS
My own product — in private beta
A multi-tenant operating system for real-estate teams, where the hard part is ingesting other people's events without corrupting your own data.
- Postgres
- pgvector
- Stripe
- WorkOS
- Twilio
- Nylas
- SendGrid
- PostHog
Context
My own product, currently in private beta with a waitlist open — which is exactly why it is the one place in this portfolio where I can show the architecture rather than describe it around the edges. Multi-tenant from the first commit, integrated with a stack of third-party providers that each have their own opinion about delivery guarantees.
The problem
A platform like this does not own most of its data. Listings arrive from a national feed. Payment state arrives from Stripe. Messages, calls and email arrive from Twilio, Nylas and SendGrid. Every one of those sources will, eventually, deliver the same event twice, deliver it out of order, or deliver it while your database is mid-migration.
The naive version works in staging and then quietly double-counts something later. Worse, the corruption is silent: nothing errors, the numbers are just wrong, and you find out weeks afterwards from someone who noticed.
What I built
- Idempotent webhook ingestion
- Every inbound event is keyed on the provider's own event ID, so redelivery is a no-op rather than a second write. This is the single highest-value decision in the system, and it has to be made before the first integration, not after the first incident.
- Anonymous-to-identified backfill
- Visitors act before they identify themselves. When they do, the earlier activity has to attach to the now-known person rather than being orphaned — which means events are stored so they can be re-attributed later.
- CREA RESO OData sync
- Continuous synchronisation against the national real-estate data standard, with reconciliation rather than blind overwrites.
- pgvector semantic search
- Semantic search over listings, in Postgres, without bolting on a second datastore to keep in sync.
- A multi-agent framework with auditing
- First-party agent orchestration with scheduled execution and an audit trail over every run — what ran, when, and what it changed.
Outcome
In private beta: demonstrable end to end, with a waitlist open and active development ongoing. It is not carrying production volume yet, and I would rather say so than imply otherwise — the architecture above is real regardless of how many tenants are on it, and it is the reference implementation for how I build everything else.
Sound like something you have?
More work
- A major Canadian airline
Booking, seat maps, flight search and automated baggage handling, integrated against airline reservation systems under a contractual throughput SLA.
- Surefire Fundraising
A Laravel platform handling catalog, checkout and transactional email for an Edmonton client whose business was built on top of it.
- What's The Deal
A Flutter app on the App Store and Google Play using location, geofencing, maps and push, with RAG-based search and a vector recommender behind it.