Skip to content
Back to LabFounder / Solo · Side project · Apartment hunting · Italy · 2024–present

CasaHunter — How I built a tool that apartment-hunts for you

AI-assisted

1,000+

listings processed daily. 73 scored HIGH (75+/100), 322 MEDIUM (50-74), 524 LOW.

€0.25

AI + infrastructure cost per user per month. 95% gross margin.

How it started

I needed to find an apartment in Modena. Like everyone else in Italy, I was checking Immobiliare.it, Idealista, Subito, and agency websites separately — each platform isolated, no way to see the full picture. The good apartments disappeared in hours. The algorithms didn't understand quality; they just sorted by recency or price.

So I built a scraper over a weekend. Parse every platform, score listings on actual quality — not keywords, not popularity, but value-for-money. Price against zone averages. Condition from photos. Commute time. Everything a human would check, but automated.

I ran it for weeks. The dashboard showed 1,029 listings across Emilia-Romagna, colored by score. Most were noise. But the high-scoring ones? They were genuinely good apartments. And I found mine using it.

The first version

The v1 dashboard was simple: a map with colored pins (green = high quality, yellow = medium, red = low), and a list view showing price, rooms, and AI score. No auth, no monetization. Just the data and the logic.

Friends started asking for access. That was the signal. Not "this is cool," but "I need this." They were doing exactly what I did — checking eight platforms manually, frustrated by the noise.

CasaHunter v1 dashboard — Map view with 1029 score-colored listings across Emilia-Romagna

Map view — score-colored pins (green = high, yellow = medium, red = low)

CasaHunter v1 dashboard — List view with AI scoring cards, 100 High scores

List view — cards with AI score, price, size and analysis

CasaHunter v1 dashboard — Statistics: 1029 listings, 73 high-scored, avg price €727, 114 zones

Statistics — price distribution, score categories, zone analysis

From tool to product

The turning point wasn't the dashboard itself. It was realizing what didn't exist in the market: nobody does AI-based quality scoring. Aggregators exist. Alerts exist. But aggregators + scoring + "is this apartment for me?" — that's a product gap.

That gap required a cost model. Running Claude Sonnet on all 1,000+ daily listings would be thousands monthly. Unsustainable. The answer: be ruthless about what deserves AI attention. Filter deterministically first (price, location, size — hard constraints), then spend AI budget on the 50% that passes the gate.

With that math solved, I decided to build a real product. Not just a bot, but a SaaS: Telegram for real-time alerts, dashboard for analysis, customer-grade infrastructure.

Three-pass scoring

Pass 1: Deterministic filter. Price, location, size, rooms. The system processes 1,029 listings per day. This pass eliminates ~51% without touching AI budget. Runs hourly, negligible cost.

Pass 2: AI scoring. Only the remaining 49% get Claude Sonnet. Real data from March 17: 73 HIGH (75+/100), 322 MEDIUM (50-74), 111 LOW. The AI reads text, analyzes photos, checks neighborhoods, compares zone prices.

Pass 3: Feedback loop. User feedback on Telegram refines scoring. Separate /outreach command auto-generates contact messages. The system learns what you actually want.

This structure keeps costs at ~€0.25/user/month. Without it, an order of magnitude higher.

Three-Pass Scoring Pipeline
1,000+ listings/day in~80% filtered at Pass 1<5 alerts/day out

The SaaS prototype

Built a Figma Make prototype to test the product vision: real-time Telegram bot alerts, persistent dashboard, saved searches, comparison tools. Not yet monetized — running on a waitlist to measure demand first.

The economics work (95% margin at €5–10/user/month), but the question remains: is demand strong enough to justify building the full SaaS? Friends asking for access is signal. Customer acquisition cost and churn are a different measurement.

Figma prototype

Full SaaS design: landing page, dashboard, onboarding flow, pricing.

Open the prototype

What the numbers say

1,000+

listings processed daily across 8 Italian platforms

73

HIGH-scored listings per day (7% signal-to-noise ratio)

€0.25

AI + infra cost per user per month. 95% gross margin at scale.

Friends asking

The earliest signal. More meaningful than metrics.

What I learned

Building intelligence costs more than building aggregation. Every SaaS tool aggregates something — the moat is what you do with the data after. For apartments, that's AI scoring that answers "is this for me?"

Unit economics determine product strategy. At €0.25/user/month, break-even is ~150 users at €10/month. That shaped everything: the architecture, the feature set, the go-to-market.

Validate first, build later. The dashboard and bot work. The SaaS is not live. I'm measuring demand before committing to customer support infrastructure. That's the hard part — not the code.

Resources

Live dashboard: https://casahunter.vercel.app/ — the scoring, the listings, the market intelligence.

Telegram bot: Running 24/7 for testing. Processes daily alerts, handles /outreach command for message generation.

These cases tell the how. For the who, there's the about page. If you want to talk, write me.

About meGet in touch