A demo, not a marketplace
An asking price is a bet on time.
This engine prices the bet.
Everything on this site — the apartment catalogue, the seller wizard, the broker dashboard — exists to demonstrate one thing: a pricing engine that tells a seller not just what their asset is worth, but what any asking price will cost or earn them, with calibrated probabilities you can check. The engine is the product. It is offered to partner platforms as an API.
Start simple
Why "what is it worth?" is the wrong question
Ask too much and the apartment sits. Every extra month costs real money: loan interest, utilities, the return your equity is not earning elsewhere — and a listing that lingers starts to look suspicious to buyers, which pushes its eventual price down.
Ask too little and it sells in a week — along with money you never see. The right price is not a point on a chart of past sales. It is a decision under uncertainty: how much time-risk are you paid to take, given your mortgage, your carrying costs, your deadline?
Portals and valuation tools answer "what is it worth?" with a single number. This engine answers the question sellers actually face: "if I ask this price, what happens — and what does it cost me?"
The engine's answer
For every price: a probability, a timeline, a euro consequence
A value model ("hedonic") estimates what the apartment is worth from its measurable facts — size, location, floor, building age, condition. This is the anchor, not the answer.
A time-to-sale model estimates, for any asking price, the probability of selling in each future week. Price 10% above the anchor? The whole probability curve shifts — and the model says by exactly how much.
A decision layer turns those probabilities into euros for this seller: who lives in the apartment, what the loan costs, how urgent the sale is. It recommends the price that maximises expected wealth — and shows the whole trade-off curve, not just its peak.
The seller sees all of this in one interactive view: a probability cloud over time, a price slider, and a live decomposition of every euro — sale proceeds, cost of waiting, carrying costs. Curious sellers can open the full mathematics, equation by equation, with their own numbers plugged in. See it live in the 60-second tour.
Going deeper
What the model cannot see is bounded, not ignored
A statistical model sees square metres, not water damage; a district, not the neighbour's drum kit. Instead of pretending otherwise, the engine gives every blind spot an explicit, bounded channel:
Facts come from the state
Type an address and the Estonian Building Registry fills the build year, floors, lift and heating; an apartment number adds its registered area, rooms and floor. If the seller's numbers differ from the registry's, the engine says so — before the buyer's bank appraiser does.
Photos are graded, then bounded
A vision model grades the listing photos — finish quality, kitchen, bathroom, light, visible red flags. A deterministic formula converts the grades into at most a ±6% shift on the value anchor: half the spread two human appraisers show on the same apartment.
Human knowledge gets a slider
Neighbours, noise, a view no photo captures — the seller shifts the anchor by up to ±5% with an explicit slider. Every layer's authority is capped, visible and auditable: model, photos, human — each bounded, none silent.
Proof, not promises
Every claim is measured — including the failures
Calibration error of the weekly sale probability. When the engine says 7%, it happens 7% of the time.
Expected extra wealth per listing vs typical seller behaviour, in a held-out decision replay.
Feature ideas measured, found wanting, and published as negative results instead of shipped.
All numbers come from a standing evaluation harness — temporal train–test split, no future leakage, decision replay on held-out listings. The honest caveat: validation so far runs against a synthetic world calibrated to real Estonian market data; the harness is the measuring stick waiting for real outcomes. The full evidence, the methodology and the API documentation live on the partner page.
Read the evidence & API docs →The demo around the engine
A full marketplace, built to exercise the API
Everything below runs on the same API a partner would integrate. The 826 active and 20 917 sold listings are synthetic — a decade of history calibrated to real Maa-amet price indices — so every workflow can be demonstrated end to end without a single real seller's data.
The 60-second tour →
No login, nothing to type: one real address walked from registry autofill to a fully interactive price recommendation. Start here.
The seller wizard →
The centrepiece. Type an address, add photos — registries and AI fill most of the form — then explore the price advisor: the probability cloud, the slider, the math.
Browse the catalogue →
The synthetic Tallinn market: filters, districts, sold archive with realised prices and days-on-market — the ground truth the engine trains against.
Market overview →
District medians, the decade-long price index and time-on-market — the aggregate view of the same synthetic world.
Broker portfolio →
The same engine at agency scale: a whole book of listings with position-vs-recommendation chips and one-click re-pricing.
Looking for partners, not listings
If you run a marketplace, an agency network or a lending desk that prices comparable assets — apartments today; the architecture generalises to cars and beyond — the engine integrates as a REST API, and every claim above comes with its measurement attached.