We partnered with NYC broker to ground our property comps API.Here's what changed across 16 markets and 43M+ transactions

#ai #mcp #claude #webdev

We run a property comparable-sales API. 16 markets, 45M+ transactions pulled straight from government records. NYC, Chicago, Washington DC, Chicago, Philadelphia, Seattle, Connectcut, Miami, Denver,Pittsburgh. International markets: UK, France, Dubai, Taipei, Dubai.

Last week we started working with an experienced NYC real estate broker — three decades on Manhattan deal desks — to ramp up our product offering and ground it in actual day-to-day broker workflow.

The first thing we did was run a sample valuation report for one of their typical downtown ZIP codes. They reviewed it, gave us thoughtful feedback, and helped us see the gap between what our data source hands us and what a broker actually uses on the phone with a client.

This post is what we built over the next 48 hours based on that collaboration.

Why "structure" isn't "property type"

NYC public sales data classifies buildings by structure: "Elevator Apartment", "Walk-up", "1-Family", "2-Family". That's how most enterprise real-estate tools bucket it too — CoStar, ATTOM, CoreLogic. It's also how every open dataset hands it over.

But a New York broker doesn't care that two buildings are both "Elevator Apartments". They care about what you actually own when you buy a unit in one of them:

Condo: you own real estate, fee simple. Standard mortgage.
Co-op: you own shares in a corporation. The corporation owns the building. You get a lease.
Condop: a hybrid that trades like a condo but governs like a co-op.
Townhouse: you own a whole house on a plot of land.
Rental building: somebody bought the whole building for $20M to collect rent. Not relevant as a comp when you're pricing a 1-bedroom.

These aren't footnotes. In Manhattan, the median co-op often trades at 50-60% of a comparable condo because of financing friction and board approval risk. A single label called "Walk-up" might cover a $700K co-op share or a $20M tenement building trading whole. Same label. Completely different transactions.

Our broker partner showed us how easily a comp set with both mixed together becomes un-defendable in a client conversation. So we rebuilt the layer that names them.

What we built across 16 markets

Ownership-aware property types

NYC now returns:

Condo · Condop · Co-op (Elevator) · Co-op (Walk-up) ·
Townhouse · Single-Family · 2-Family · 3-Family ·
Rental Building · Mixed Use

Same underlying NYC DOF building-class codes, mapped into the 10 categories that a broker actually distinguishes in day-to-day valuation work.

Unit sales vs whole-building sales

A 6-unit rental building selling for $15M and a 1-bedroom co-op selling for $800K are both "transactions" in raw data. They shouldn't be in the same comp set.

Added a sale_type=unit | whole_building | all parameter — defaults to unit. Multi-family rental buildings, mobile home park bulk sales, commercial condos: all filtered out of default residential comps. Still queryable when you explicitly want them.

NYC and Miami both publish building-size data, so both got this filter natively. Paris, Taipei, and Dubai got the equivalent through residential-only defaults (more on that below).

Per-city re-classifications

Each market has its own quirks, driven by how its public records are structured:

Philadelphia: 60% of residential sales are rowhouses. The ingest was calling them all "Single Family" (none were even labeled "Townhouse" in the database). We reclassified 160,000 rowhouses and recovered 14,000 condos that had been hidden in the Single Family bucket.
Seattle: only 34 condos in 803,000 transactions. Because King County Assessor stores condos in a separate data extract that our ingest never joined. Added the join. 223,000 condos recovered.
Washington DC: 18,000 transactions labeled generically as "Residential" were actually rowhouse-to-apartment conversions. Split into "2-4 Family" and "Apartment Building".
Paris: 36% of the French DVF dataset is "Dépendance" — garages, sheds, parking spaces. Not dwellings, but they were polluting residential comps. Added a residential-only default filter.
Taipei: 22% of rows were labeled "其他 / Other". These turned out to be land sales (addresses ending in 地號, the Taiwanese land parcel code). Reclassified to "土地 / Land" and excluded from default residential comps.
Dubai: DLD records 16,000 whole-building transfers plus Land, Office, Hotel Apartment, Hotel Room, Shop, Warehouse. All excluded from default residential results.

Statistical honesty in trend charts

A per-type price trend chart built from 2 sales per month is noise, not a trend. Every chart now requires at least 60 sales and 6 months of data before it renders. Thin-volume types still appear in the summary tables (their medians are valid data points) but they don't get a chart line that would mislead a reader into thinking "Townhouse down 56%" when it's really just three high-end sales in one month.

It sounds obvious. A surprising number of dashboards ship without this.

Default language matters

The French pdf report is in French by default. Methodology is in French. Even the database values carry the accented characters because the source DVF data does.

The Taipei report is in Traditional Chinese by default. NotoSansTC font embedded. No emoji (they fail silently when the font doesn't support them — we use [注意] instead of ⚠).

Singapore is bilingual with a Chinese(zh) toggle. Dubai is English with an AR toggle.

When we added new broker-friendly labels, we had to add each translation everywhere. "Co-op (Elevator)" in English becomes the corresponding Chinese term in Taipei. We kept tripping over this — leak one English label into the French methodology block and brokers notice instantly.

What this costs compared to the alternatives

We're not claiming parity with the enterprise tools. CoStar has private databases, building photos, tenant rosters, leasing data. Syndicating a $200M industrial portfolio? You're still using CoStar.

But if you're a broker running 5-10 comps searches a week, pricing listings, grounding buyer offers, or generating seller-facing valuation reports — at $29/month, it costs less than the coffee you bought at the closing table.

Features the cheap tools don't have either

Built because brokers asked:

Per-type rental yield analysis: yield estimates broken down by property type. A Chicago report shows Condo 3.6% vs Co-op 4.0% vs Townhouse 7.8%, with a methodology paragraph explaining the HUD SAFMR rent source and the expense-ratio assumptions used.
Type Premium Analysis: plain-English narrative: "Condo median is 116% higher than Co-op ($1,995,000 vs $923,719)." Reads like something you'd actually tell a client.
Supply & Demand Signals: recent 3 months vs prior 3 months, plus a narrative: "Falling volume + rising prices: limited supply supporting prices."
Size-Adjusted Valuation: per-square-foot median from similar-sized comps. Not a raw median that ignores whether the subject is 800 or 2,800 sqft.
Price Sensitivity Analysis: three discount scenarios (5%, 10%, 15%) with estimated speed-to-sale boost. Useful when a seller is anchored 10% too high.
MCP server for AI agents: Claude, ChatGPT, and similar AI agents can query property comps as a tool. A broker can ask "what did Upper East Side 2-bedrooms trade for in the last 90 days" in plain English and get real transactions back.
Proper API:OpenAPI 3.1 spec, RapidAPI distribution, auth, rate limits, CSV export. Not a scraped HTML table.
Languages: English by default, French, Traditional Chinese, Singapore bilingual, Arabic toggle on Dubai. Because international investors are a real segment.
We welcome Proptech developers/real estate agencies to partnership and or tailor solutions