This case study documents an ongoing project. The product is still in development and not yet public.
← All work
AI Product · UX/UI · Community Platform · 2026

A healing marketplace
being built with AI from zero

Designing and building a full-stack Hebrew platform that connects seekers with therapists in Pardes Hanna. Smart matching, 4-step signup, admin verification. Built with Claude and Lovable, still in progress.

Role Founder + Art Director of AI + Product Designer
Year 2026
Stack React · Supabase · Lovable · Claude
Status In development, not yet public
10+
Pages built
4
Signup steps
1
AI edge function
Zero
Previous codebase

A healing community with no digital home

Pardes Hanna has a quiet but rich ecosystem of therapists, facilitators, and healers. Bodyworkers, shamanic guides, therapeutic writers, astrologers, group retreat leaders. They all existed in silos: word of mouth, scattered Instagram accounts, Facebook groups nobody searched.

The request was clear. Build a platform where people looking for healing can describe what they need, and the platform finds the right match for them. Not a rigid filter form. Something that understands human language and the nuance of emotional need.

The platform had to work in Hebrew, support RTL layout throughout, and feel warm, not clinical. A marketplace with soul, not a directory with a logo.

"We wanted a place where someone could type 'I'm going through a divorce and need support' and actually get matched to a real person who can help, not a list of checkboxes."

Client brief, January 2026

Design system first, then architecture

Before writing a single component, I defined a visual language rooted in the identity of the place. Pardes Hanna is green, warm, earthy. The design system reflects that: soft sage green as the primary tone, warm earth tones for secondary surfaces, and a gentle cream background that reads like paper, not screen.

The key tension in the design was clinical vs. communal. Marketplace UX patterns tend to feel transactional. A list of providers with ratings and prices. That framing would have killed the product. The solution was to lead with language, not profiles. The SmartMatchingForm is the homepage centerpiece, not a search bar or a grid of thumbnails.

Without HaMerhav
Search Facebook groups for recommendations
Find a name, no bio, no specialization context
DM on Instagram and wait for a reply
Unsure if this person does what you actually need
Start over with the next name on the list
With HaMerhav
Type what you're going through in natural Hebrew
AI matches your words to verified provider profiles
See specializations, services, and prices upfront
Contact via WhatsApp directly from the profile
Provider is verified. Trust is built into the system.

RTL is not a CSS flag. It's a worldview.

Hebrew RTL layout changes more than text direction. It changes spatial logic. In LTR interfaces, primary actions sit on the right. In RTL, they shift left. Progress flows right to left. The eye enters from the opposite edge. Every component had to be designed with that in mind from the start.

I applied dir="rtl" at the page root and used Tailwind's RTL utilities throughout. Navigation reversed. Form labels aligned right. Service cards stacked naturally. What felt effortless to Hebrew users was a deliberate architectural decision made at line one.

Design principle

Warm, not clinical. The palette is sage green, not blue. Cards have soft shadow and rounded corners. Copy says "ספרו לנו מה עובר עליכם" (tell us what you're going through) instead of "submit a query".

Three systems that make it work

AI Smart Matching
Users type a natural-language description of what they need. A Supabase edge function sends the query to an AI model, which returns ranked provider matches with a reason for each recommendation. No filters. No categories. Just language.
🤵
4-Step Provider Signup
Therapists and facilitators go through a structured onboarding: authentication, public profile, professional specializations, and certificate uploads. Each step is validated with Zod. Progress is preserved across sessions.
🔒
Admin Verification Panel
A protected admin route lets the platform owner review pending providers, view uploaded certificates, expand profiles inline, and verify or reject with a single click. Status updates via a Supabase edge function with RLS enforcement.
📋
Rich Provider Profiles
Each provider has a full-page profile with bio, specialization tags, service listings by type (1:1, workshop, group, retreat), pricing, duration, and a direct WhatsApp CTA. Verified badge displayed prominently.
Therapist signup: 4 steps
1
Authentication
Email signup or social login via Supabase Auth
2
Public Profile
Name, title, location, online availability, profile photo upload
3
Professional Details
Bio (500 chars), specialization tags, price range, WhatsApp, Instagram
4
Certificates
Upload credential files to Supabase Storage. Pending admin review.

Full stack, zero infrastructure ops

React 18 + Vite 5
Frontend framework with fast dev server and instant HMR.
🎨
Tailwind CSS
Utility-first styling with a custom healing color system and RTL utilities.
🗃
Supabase
Postgres DB, Auth, Storage, and Edge Functions. Entire backend in one place.
🤖
AI Matching Edge Fn
Supabase edge function that sends user queries to an AI model and returns structured matches.
🧰
Lovable
AI-native dev environment. Translated design decisions into working React instantly.
🔒
TypeScript 5 + Zod
Type-safe schemas throughout. Supabase-generated DB types, Zod for form validation.

Designing and building simultaneously

The build process was non-linear. I designed components by describing them in natural language, then reviewed the generated code, adjusted visual decisions, and iterated. Lovable handled the scaffolding, routing, and component wiring. I handled the decisions about what the product should feel like and why.

There was no Figma mockup stage. The design lived in the conversation. The RTL system, the color palette, the messaging tone were all defined through prompting. Lovable's ability to maintain state across a complex codebase meant I could focus on product thinking instead of file management.

The AI matching edge function was the most complex piece. It required thinking through what a quality recommendation looks like, how to format the prompt so the AI returns structured JSON, and how to handle cases where no match is strong enough to surface. That judgment call was mine. The implementation was Claude's.

"I didn't write the code. I made every decision about what the code should do, how it should feel, and why. That's a new kind of authorship."

Merav Shaked

A real product, delivered from zero

HaMerhav shipped as a working platform. Providers can sign up, create profiles with services and certificates, and wait for admin verification. Seekers can describe what they need and receive AI-matched recommendations. The admin panel allows the platform owner to manage the entire provider registry from a single interface.

What took this from experiment to product was the clarity of design intent. Every component had a purpose. The warm palette communicates safety. The natural-language search removes friction. The verification flow builds institutional trust in a domain that runs entirely on personal trust.

Hebrew RTL AI Matching Full-Stack Product Multi-Step Forms Admin Panel Supabase + RLS Healing Design System Community Platform Lovable + Claude
Next project
Building meravish.com with AI