← Work / 01 / 2023 → Now
RoomIQ.
Market intelligence for room-rental investors. 600+ US cities, refreshed daily.
The product
Room-by-room rentals (co-living) are one of the highest-yield corners of US real estate, and one of the most opaque. There’s no MLS for rooms. Investors underwrite six-figure decisions on gut feel and stale spreadsheets.
RoomIQ fixes that: track occupancy, pricing, and revenue across hundreds of US cities, analyze any property down to the room level, and forecast returns before you invest. Search a market, read its trends, rank its properties, then run an AI-powered revenue forecast on any address.
The data platform
This started as a freelance data contract and grew into a company; I’m now CTO. The backbone is a Dagster-orchestrated pipeline: 44 assets across 21 scheduled jobs, ingesting 80,000+ listings daily into a star-schema PostgreSQL warehouse covering 116 metros, with freshness sensors, partition-level quality gates, and a 2,500-line scrape-quality reconciliation ledger that audits raw JSON against the warehouse every day.
The ingestion side uses request-budgeted, fingerprint-cloaked scrapers (raw-first, normalize-later), a Cloudflare R2 media mirror, and even FEMA flood-risk overlays as a property signal. Roughly 60,000 JSON documents parse into the warehouse in about 12 seconds across 8 parallel workers.
The ML
The production model is deliberately unglamorous, and that is the point.
- Room-rent prediction: an XGBoost regressor over 63 leakage-controlled features (target-encoded categoricals, K-Means geo-clusters, market aggregates computed from the training split only), hitting 8.2% MAPE on held-out room-level rent. I prototyped fancier ensembles reaching for foundation models and retrieval, and they were worse, so the boring, accurate, explainable model is the one that ships.
- Market forecasting: weekly forward-looking forecasts of stock, price, occupancy, and revenue per city, with uncertainty bands and enforced metric identities. A fill-time model estimates 7/14/30-day fill probability and sweeps candidate prices to generate pricing guidance.
The product surface
An 88-route FastAPI backend (Redis, SuperTokens auth, Stripe billing, rate limiting) serves a Next.js 15 dashboard with 260+ components: market KPIs, trend charts with private/shared bath breakdowns, metric-driven property ranking, and Leaflet vector maps with occupancy and revenue-coded clustering.
The methodology stance
RoomIQ’s differentiator is discipline, not dazzle: consistent filter scope across every tab, explicit date windows, and metrics that map one-to-one to dashboard fields. Every number is explainable, because people commit real capital against it. We even measured our own systematic bias (a 3.8% underestimation) and shipped an explicit calibration correction.
Outcome
RoomIQ is live at roomiq.io with free and Pro tiers, tracking 6,600+ active properties. What began as “can you scrape some listings?” is now the analytics layer an entire niche of real-estate investing was missing.