---
date: 2019-09-01
type: ship
title: belowmrkt
slug: belowmrkt
project: Independent
kicker: Adaptive property search that learned from how people browsed.
excerpt: Designed a machine-learning property-search system where filters re-shaped themselves from quick scans, shortlists, and revisits. Intelligence stayed invisible so users could focus on discovery.
cover: /assets/covers/hero-belowmrkt.webp
palette:
  accent: "#6B8294"
  source:
    brand: "Pantone"
    name: "5415 Slate"
  role: Lead Product Designer
  pull: Filters that adjust themselves taught us how people actually shop for a home.
tags: [ai, real-estate, machine-learning, 0-to-1]
---

belowmrkt was an adaptive property-search product. Instead of asking users to set filters and re-set them on every visit, the system learned from quick scans, shortlists, and revisits, refining results in real time without ever showing the user a slider.

## Context

We had weeks, not months. Almost no training data. The model had to start learning from the very first session, and the design had to prove that intelligence without breaking the simple act of browsing listings.

![belowmrkt screen system shot from above on a dark grey backdrop. A dozen iPhones lie at angles across an isometric grid, each showing a different surface of the app: the property contact form with a red "Send Message" CTA, the listing detail with map and "Get Directions", the Florida home feed with "Ocean Club Residencies", the Sort & Filter modal, the Create Alert form with category chips and price slider, the saved-searches Alerts list, the Settings page, and the search results with property cards.](/assets/projects/belowmrkt/bm-overview.webp)

## What I designed

The interaction model around micro-patterns: which cards a user paused on, which they shortlisted, which they came back to. Each became input for the model. The interface stayed minimal by design. Every gesture carried weight, every scroll improved the next result.

![Four iPhones in a row on a flat red backdrop, each showing a deeper scroll into the same Sort & Filter modal. From left to right: the entry state with Buy/Rent toggle, Sort by, Categories (Single Family Home, Half Duplex, Condo, Full Duplex, Triplex, Fourplex), Location and a Bedrooms stepper at 2; a "Show Sort Options" panel revealing Newest First, Oldest First, Nearest To Me, Lowest Price, Highest Price, Lowest Price Per Sqft; an islands and areas chip grid (Andros, Bimini, Abaco, Long Island, Eleuthera, Bay Street, Eastern Road, Adelaide Road); and a Categories carousel with Price Range and Per Sqft Range sliders. A red "View 45 Units" CTA pins the bottom of every screen.](/assets/projects/belowmrkt/bm-filters.webp)

## The decision that shaped it

Hide the intelligence. The category convention was to expose the AI: confidence scores, reasoning panels, "we recommended this because…" explanations. That language reads as marketing for the model and noise to the user. We chose to make the system feel attentive but invisible: subtle layout shifts and result re-ordering instead of explanatory chrome. The trade-off was losing the "look how smart we are" pitch in exchange for an interface that felt intuitive rather than algorithmic.

## What it left behind

Launched fast and gained early traction. Users spent less time adjusting filters and more time exploring listings. The pattern (letting a small model learn from interaction shape rather than explicit input) held up far better than its training data deserved.

![Three iPhones diagonally cascade down a soft grey backdrop. Top phone: the Create Alert sheet with keywords input, Categories chips, Location button, Bedrooms / Bathrooms steppers (2 each), a Price Range slider and a red "Create Alert" CTA. Middle phone: a "32 Results Found" search returned for "2 Bedroom City Center" with property cards (Skyline Drive, 25 Polnciana Cay) at $900,000 each. Bottom phone: the saved Alerts list showing "2 Bedroom City Center: Within 2 Miles…" plus other saved searches with property counts.](/assets/projects/belowmrkt/bm-alerts.webp)

## Role

Lead Product Designer. Worked side-by-side with engineering on what the model could meaningfully respond to.