Case studies

Enter the password to view this page.

Access granted

Esther Yang
Work About Resume Notes Contact

← Work

01

Trust Infrastructure for a Two-Sided Marketplace

Designing the system behind how 550K+ professionals discover, evaluate, and build relationships.

Two-Sided Marketplace Relationship Systems Trust

The platform connected loan officers and real estate agents, but the product had no consistent model for how those relationships were discovered, evaluated, formed, or managed.

I designed the relationship model for the referral system and Agent Insights, then applied the same states and actions across discovery, evaluation, outreach, pairing, and ongoing management.

Problem Model Surfaces Agent Insights Pairing Leverage Outcomes Closing
Role
Lead Product Designer
Scope
Designed the relationship model for the referral system and Agent Insights, and carried it through discovery, evaluation, outreach, and LO–agent pairing.
Team
2 Product Designers · Product Manager · Engineering
Ownership: Sole designer for the referral system and Agent Insights
Scale
400K+ real estate agents · 150K+ loan officers
Outcome
81 NPS · 64% cold-invite click-through · Company acquired

Confidentiality Notice: This case study has been anonymized. Visual details, content, and identifiers have been modified to protect confidential information while preserving the underlying design challenges and decisions.

01 — Problem

The problem wasn't a referral flow. It was the relationship model underneath it.

A two-sided professional network: loan officers and real estate agents needed each other, with different goals, context, and risk.

Pairing was also a core growth mechanism for the business. Each new relationship expanded the network and increased the value of the platform for both sides. But the product couldn't optimize for growth at the expense of trust: an agent could only be paired with one loan officer, so forming a relationship was consequential.

People needed enough information to evaluate someone first. The product also had to show pending vs. active states and stay within U.S. consumer-protection requirements, including RESPA.

What was missing: one consistent path from discovery → evaluation → outreach → pairing → ongoing management. Warm and cold paths were treated as separate problems; they needed the same underlying structure.

Fig. 01 — Relationship network loop

  1. Discover

    Transaction data + Agent Insights surface relevant agents and LOs

  2. Initiate

    Pairing + outreach

  3. Connect

    Agent + LO relationship

  4. Distribute

    Co-branded marketing reaches new users

Discovery, outreach, and pairing only work as one loop if they share a relationship model.

02 — Model

The work started with the relationship model.

Design question

What does someone need to know and see before deciding to form a professional relationship?

Decision

A shared structure for discovery, evaluation, outreach, pairing, and ongoing management, used for both warm and cold paths.

Evaluate

Enough signal before contact

Transaction history, production, price band, loan types, and market activity, without treating a cold introduction as if trust already existed.

Represent

People, states, constraints

How a potential relationship appears, what changes before vs. after an invite, and how pending vs. active reads when only one LO pairing can be active.

Relationship rules

Context determines the path

An agent could pair with one LO, while an LO could pair with multiple agents. When pairing wasn’t possible, Follow kept the connection available.

After pairing

Keep managing the relationship

Dashboard activity, ongoing suggested partners, and clear rules for how the platform speaks about someone in outreach.

Fig. 02 — Relationship stages

  1. Discover
  2. →
  3. Evaluate
  4. →
  5. Outreach
  6. →
  7. Pair
  8. →
  9. Manage relationship activity
The stages stay consistent across warm and cold relationships; what changes is the information and action available at each stage.

03 — Surfaces

I applied the same stages across the product.

Discover, evaluate, outreach, pair, and manage weren’t separate feature logics. I used the same relationship states and actions in discovery, Insights, outreach, pairing, onboarding, and the home dashboard.

  1. 01 Discover Find relevant agents and LOs
    Agents directory filtered by zip codes, with share of wallet metrics for each agent
  2. 02 Evaluate Read production and market signals
    Agent profile overview with top production highlight, share of wallet, and buy-side versus list-side transactions
  3. 03 Outreach Follow or initiate contact
    Optional invite message draft with send to self, copy invite link, and send invite actions
  4. 04 Pair Establish the professional relationship
    Paired confirmation showing two profiles connected with co-branded marketing benefit
  5. 05 Manage Ongoing dashboard suggestions
    My Agents dashboard with suggested agents setup card, pending invites, and paired or pending status list

04 — Agent Insights

Agent Insights became the evaluation layer of the model.

LOs could use Agent Insights to research both agents and other LOs before taking action.

Fig. 03 — Discover → Evaluate → Outreach

From discovery to evaluation to outreach.

Agent Insights surfaced transaction history, production volume, price range, loan types, and listing maps. LOs could follow an agent without initiating a pairing, keeping the relationship lightweight until they had enough context to make contact.

Design question

What does an LO need to know before deciding to pursue a relationship?

Decision

Put evaluation before outreach: give LOs a nationwide agent database with the signals that make contact credible.

05 — Pairing

Pairing made the relationship rules visible in the product.

Pairing had to answer the same questions on both sides: what you know before inviting, how invite state is shown, and what changes when someone accepts.

Warm path · existing relationships

Search alone asked LOs to remember who they'd worked with. With transaction data, I surfaced suggested agents under search so the product could help them recall and evaluate existing relationships before inviting.

The relationship itself was asymmetric: an agent could be paired with only one LO, while an LO could pair with multiple agents. If an agent was already paired, another LO couldn’t invite them, but they could still Follow. That exclusivity was a relationship rule the product had to surface clearly.

Fig. 04 — Pair

Invite agents to pair modal showing agents you've worked with, shared past transactions, copy invite link, and Invite 3 agents
Warm invite: suggested agents from past transactions, or a copyable link.

Relationship states

The relationship state determined what the LO could do next.

For existing Platform B users, accepting the invitation established the pairing. For non-users, the invitation remained pending until they joined and accepted.

  1. Unpaired

    Agent profile card with Follow and Invite actions available
    • No existing relationship
    • Pairing not started

    Action Follow · Invite

  2. →
  3. Invite pending

    Agent profile card with Follow available and invitation pending status
    • Invite sent; not paired yet
    • Pending until they join and accept

    Action Follow

  4. →
  5. Paired

    Agent profile card showing an established Paired relationship state
    • Relationship established
    • Recipient is on Platform B

    Action Follow

Constraint · not a linear next step

Already paired with another LO

Agent profile card with Follow available and Already paired action disabled
  • Agent can only pair with one LO
  • Invite unavailable for this LO

Action Follow

Trust · ownership at outreach

Design question

Was invite drop-off a writing problem—or a trust problem?

Decision

Make the message transparent and give users control.

Early invites used standardized platform messaging to reduce the effort of sending invitations. We saw drop-off at the invite modal, and Customer Support heard from LOs and agents who were uncomfortable: they didn’t know exactly what the email would look like, or what would be sent in their name.

I redesigned the invite experience around that feedback. Convenience still mattered, so preview wasn’t required on every send. Users could see the outgoing message, optionally edit it, and add a note in their own voice. The platform could facilitate and draft, but it shouldn’t speak for the LO without their knowledge.

Constraint

Pairing exclusivity reads as relationship state.

Already-paired agents stay reachable through Follow, while unpaired agents can be invited. The product makes the relationship state and its available actions explicit.

Outreach

Don’t speak for the user

Available preview before send, plus an optional note in the LO’s own voice. Drop-off and support feedback pointed to a trust problem as much as a convenience problem.

Evidence Invite-modal drop-off and CS complaints about unknown outbound email.

Cold path · no shared history

Without shared history, the same pairing rules still had to establish trust. Two send paths: platform-sent email with an available preview and optional self-send, or a unique invite link from the LO’s own inbox. Click → automatic pairing when the recipient was already on Platform B; otherwise the invite stayed pending until they joined and accepted.

Smart email later helped draft the message; the LO still owned the send. Cost stayed explicit: LOs pay; agents don’t.

Fig. 05 — Manage

Side-by-side LO and agent onboarding screens for setting up a pipeline and pairing with a loan officer, with annotations on suggested partners and pairing constraints
Same pairing pattern continues into onboarding: suggested partners on both sides, with the exclusivity constraint made explicit.

06 — Leverage

The referral flow was visible. The relationship model was the lasting part.

One relationship model

↓

  • Discovery
  • Outreach
  • Pairing
  • Relationship management

The referral experience was high-stakes and easy to point to, but it was one implementation of the relationship rules underneath, not the system itself.

Users didn't have to think about the rules. They just had to trust us with the relationship.

07 — Outcomes

What held at network scale.

Network scale

  • 550K users across the network
  • 400K+ agents · 150K+ LOs
  • 81 NPS · Company acquired

Relationship behavior

  • 59.5% LO → Agent invite CTR
  • 64% cold-invite CTR

These figures describe the network the model operated in. They are not outcomes attributed solely to this work. Invite cohorts showed people acting on introductions the product facilitated. Co-branded marketing also helped existing relationships bring new users into the network.

Fig. 06

LO → Agent invite funnel

30-day cohort

Unique users

Click-through rate

59.5%

6000500040003000200010000
5,000
Invite Submitted
4,700
Email Sent
4,200
Email Opened
2,500
Email Clicked

Fig. 07

Cold Invites

30-day cohort

Unique users

Click-through rate

64%

10008006004002000
840
Invite Submitted
820
Email Sent
750
Email Opened
480
Email Clicked

Funnel evidence from 30-day invite cohorts: what progressed after invitations were sent.

08 — Closing

The outcome was a common model for how relationships worked.

Once those rules existed, new relationship experiences could build on the same states and actions instead of inventing one-off flows.

What I'd carry forward

Defining relationship states and constraints earlier would have made it clearer what belonged in each experience as the product grew.

← Previous Less AI, Doing More Next → Turning “Is it working?” into “Who should I call?”

Ready to build something  Ready to build something lasting?

hello@estheryang.is

LinkedIn