← Back
Case study·Co-founder & product designer·Oramo·2026–present

How do you put an AI in a product it has no right to see?

Oramo is the property management platform I’m building for Romanian landlords in a three-founder bootstrap. Two features I led from design through implementation — with AI (Claude Code) as a force-multiplier, not a replacement — and the question about data, trust, and multi-tenancy they turned out to share.

Liveoramo.ai · app.oramo.ai
RoleCo-founder & product designer · design and implementation lead in a three-founder team · Oramo (Jan 2026–present, launched April 2026)
Team3 founders · real-estate, financial strategy, product & engineering
StackNext.js · TypeScript · Supabase · Vercel · Claude Code
DeliverablesEnd-to-end product design, AI integrations, data architecture, implementation, launch
Outcome
  • Launched April 2026 into a 0.00B rental market with near-zero software adoption
  • Property onboarding under three minutes
  • Market-intelligence engine covers 0+ Romanian neighborhood zones from day one
  • In-product AI assistant answers portfolio questions without learning its way to anyone else’s

Context

Tens of discussions across the country — 30+ with landlords managing portfolios of 1 to 10+ properties, 10+ with agencies — plus competitive review and structured survey work surfaced one finding: most landlords operated their portfolios with zero profitability visibility. The tool of record was Excel paired with WhatsApp. Supply-side software penetration was near zero.

My role

Co-founder and product designer. In a three-person founding team — real-estate expertise, financial strategy, and me on product strategy, design, and implementation — I own the product surface end-to-end. Bootstrapping meant leveraging AI (Claude Code) heavily as a tool to extend what three founders could ship — not to replace roles on the team. AI as force-multiplier, not solution.

Three decisions that mattered

1. Build the data layer as if it were the product. The market-intelligence engine parses asking-price data from three of Romania’s largest real estate listing websites into per-zone benchmarks. The move that matters is what comes next: as Oramo users sign leases inside the product, the platform accumulates the first real-rental-price dataset in Romania — actual closed-lease data, per zone. The listing scrape is designed as a bootstrap; the surface stays stable while the data behind it gets better.

TodayBootstrap: listing dataWebsite 1Website 2Website 3Parse & normalize750+ zonesPer-zone benchmarkLandlord-facing surfaceAsking prices. Accurate enough to ship.Tomorrow →Replace the source; keep the surface.Signed leases inside OramoPer-zone benchmarkactual prices
Fig. 1The surface stays stable; what sits behind it is designed to be replaced.

2. An AI assistant inside a multi-tenant product is a permissions problem in a conversation costume. The in-product assistant answers a landlord’s questions from context they legitimately own. The hard decision was what it must never do — reveal anything about another landlord’s portfolio, a neighboring property’s rent, or any global detail — on every turn, for every phrasing.

Working principle: the assistant’s context is the user’s own data plus anonymized zone-level aggregates. Anything more specific than a zone is outside the boundary, regardless of how the question is phrased. Implementing that — retrieval scope, refusal behavior, confidence of refusal — is most of the work.

What the assistant knowsUser’s own data + anonymized aggregatesThis landlord’s portfolio (all of it)Zone-level anonymized benchmarksTenant & lease data — their leases onlyOutside the boundaryRefused, regardless of phrasingAnother landlord’s dataNeighboring property’s rentAggregate stats about Oramo itselfExampleSame tool, two questions, two outcomes“Which of my flats is underperforming?”→ Answers from the landlord’s portfolio“What does my neighbor’s flat rent for?”→ Refuses; offers the zone benchmark
Fig. 2The design work is making refusal feel helpful rather than defensive.

3. Choose conversation where the alternative is a form nobody wants to fill in. Small-portfolio landlords aren’t power users. Dashboards reward people who want to look at their data; conversation rewards people who have a question. Nobody we spoke to said “I wish I had a dashboard.” They said “is this one still worth it?”, “am I leaving money on the table?”, “did Gigi pay?” The interface followed the question.

What shipped

What I take from it

The most consequential decisions on an AI-native product are about what the system should and should not be able to know. Those decisions live in the data layer, not the interface.

And: a small team using AI as a tool, not a solution, can now ship what would have taken a much larger team two years ago. That’s the shape of work I’d bring into a larger team — someone who designs the surface, understands the system underneath, and makes the tradeoff calls in the room alongside engineering, not in isolation.