Melkyar — Real Estate Matching Engine
Agencies in Iran trade listings as free-form Telegram posts full of typos and shorthand. Melkyar pulls those posts from an agency's public channel, extracts area, neighbourhood, rooms, floor, price and amenities into a structured file, and scores every new file against every active buyer. When a file clears the agency's threshold, the buyer is notified over the agency's own Telegram bot. It is a private workspace per agency, not a public marketplace.
// Architecture
A FastAPI service on the shared Orv SaaS core with a Next.js 14 panel and PostgreSQL 16, one tenant per agency. A background sync reads the public t.me/s preview of each registered channel every few minutes, de-duplicates posts by fingerprint, and passes the text to an LLM for extraction, with a rule-based Persian parser as a fallback so the product still works with no API key. Matching itself is plain, deterministic code: no model decides who sees which file.
// Key capabilities
Explainable scoring
Each buyer–file pair gets a 0–100 score from weighted criteria (price 30, neighbourhood 25, rooms 15, area 15, amenities 10, year built 5), and every match stores the reasons that produced it.
Buyers search in plain Persian
On the agency's bot a buyer types something like “three bedrooms in Saadat Abad with parking”; it is parsed into criteria, saved as a lead, and answered with the closest files.
Map-aware neighbourhoods
With Neshan geocoding, files and wanted neighbourhoods get coordinates, so a file just outside a buyer's area can still score instead of missing on spelling.
Tools for the agent
The panel drafts ad copy for a file and suggests a price from the median price per square metre of comparable files in the agency's own data.
// Technology stack
- Python / FastAPI
- Next.js
- PostgreSQL
- LLM extraction
- Neshan Maps API
- Docker
// Inside the project
// Next step
Need something like Melkyar — Real Estate Matching Engine?
Tell us what you're building. We'll reply with how we'd approach it, a rough timeline and the first step.
Start a similar project