PROJECT

Deal Analyzer: Helping Lenders Decide 8X Faster

Working with expert financial users to get them from pre-screen to closed deal, 50% faster.

Fintech

AI Product

B2B SaaS

0 to 1

cliff notes

Overview: Turns a broker’s deal package into a cited go/no-go pre-screen.


Challenge: Offshore hand-keying added 24 to 48 hours to every deal, and underwriters did not trust an AI that led with a verdict.


Solution: Data first, AI recommendation second. Every value cites its source, and missing items stay visible.


Impact: From 24 to 48 hours to minutes. Live with our design partner, and sales leads with it.

Deal Analyzer deal overview: deal structure, property basis and sources and uses tables beside a checklist of what is still missing. Sample data.

DATE

June-September 2026

MY ROLE

Product Designer

TOOLS AND TECH

  • Figma

  • FigJam

  • Claude AI

  • Linear

CORE TEAM

1 Product Designer

1 Product Manager

1 Front End Engineer
1 Full Stack Engineer

Project Impact

Speed to decision

0

x

faster

Accuracy

0

%

error reduction

Provenance

0

x

more source-linked analysis

My Scope

I led design across the whole arc, from the first discovery calls to the screens engineering is building now. I set the product’s information architecture, ran design reviews directly with the CEO and Head of Engineering, sat in the customer validation sessions with the underwriter myself, and prepped the work for handoff with a versioned Figma archive and annotations. Because this is the product leadership sees as most sellable, the design decisions sat unusually close to revenue, and the work has since pulled me into a cross-functional role shaping platform patterns for the whole app, not just this surface.


Read more about Deal Analyzer on Hypha

step 1: discovery

Speed, without handing over the judgment

Every broker opportunity arrived by email, went to an offshore team to be hand-keyed, and only then reached the underwriter, so every deal started 24 to 48 hours behind. Speed to a yes or no was the pain.

Our design partner’s pre-screen, as their team described it in discovery, next to the same job in Deal Analyzer. The broker email still starts it, and the underwriter still makes the call.

“Holy shit, we don’t need to send this to offshore.”

Lead Underwriter, our design partner

Step 2: define the problem

User Problem

Every broker deal at our design partner waited 24 to 48 hours for an offshore team to manually type in data before an underwriter could look at it, while the sponsor kept shopping it elsewhere. Speeding up that step wasn’t enough on its own. The underwriters didn’t want an AI verdict; they wanted numbers they could check for themselves before making the call.

How might we…

How might we help underwriters reach a go/no-go on a new deal in minutes, while keeping the credit call in their hands?

step 3: who it's for

Two users, one shared deal

Both roles come from our design partner’s origination team. They work the same deal from different vantage points.

Lead Underwriter

The Deal Lens

GOALS

  • Give the broker a yes or no on a new deal the day it arrives

  • Check any number against its source before deciding

  • Know what a deal still needs, and ask the broker once

PAIN POINTS

  • Every deal waited 24 to 48 hours for an offshore team to hand-key it

  • Won’t act on a broker’s figure or an AI verdict without seeing the source

  • Tracked the diligence checklist by hand in a spreadsheet, including fields that didn’t apply to the deal type

Head of Origination

The Portfolio Lens

GOALS

  • See where every deal stands, by stage and age, at a glance

  • Put the team’s time on the deals worth pursuing

  • Know who is working which deals

PAIN POINTS

  • More deal packages arrive than the team can work, and good ones go to faster lenders

  • Deal status lived in the CRM while the analysis lived in spreadsheets

  • Needs a summary to make the call, not every extracted field

step 4: key design decisions

Turning the finding into the screens

Turning the finding into the screens

V1: A quick, low stakes prototype

V2: Designing for the expert’s judgment, not around it

THE PROBLEM

V1 led with the AI verdict. Our customer wanted the opposite: the numbers, dense, so he could judge.

THE SOLUTION

Extracted numbers first in scannable tables; recommendation and buy box present but de-emphasized; provenance on every value.

Main learning from user validation:

Expert users do not trust AI to do their work for them. They want to look at the full picture themselves, and keep the judgment layer. Designing AI products for expert users involves learning where AI tools are the most useful, not forcing them into every experience.

V3: The checklist: from a side panel to a card in the page

THE PROBLEM

Users asked for a constant read on what’s answered vs. missing, but engineering pushed back on a new checklist and wanted to reuse the attributes page.

THE SOLUTION

I met engineering halfway with a card built on the existing attributes table, then worked with my PM to keep the side panel on the roadmap and kept pushing for it.

step 5: leading the work

Running the reviews, not just presenting in them

Executive reviews with the CEO and Head of Engineering are easy to lose to tangents, so I ran them like working sessions. The numbered steps below are exactly that approach:

1

Sent a pre-read ahead of each review so the room arrived aligned instead of reacting live.

2

Walked in with the decisions I needed to leave with, and an argument for each tied to a user’s words.

3

Led the conversation through those decisions, showing the screens in service of them, not as a tour.

4

Kept the room on task so we resolved things and kept moving, rather than reopening settled ground.

To keep the build matching the intent, I ran a handoff system alongside the reviews:

1

Annotation system in Figma documenting interactions, states, and edge cases on each frame.

2

Versioned archive (in-progress, ready, archived) so the team always knew status.

3

Weekly design/eng syncs comparing the live build against Figma, filing regressions.

step 6: impact and what it spun off

step 6: impact and what it spun off

One product that seeded platform-wide patterns

End-to-end workflow built and validated with the customer; Kanban shipping to production; the topology deployed to our customer, replacing their old pre-screen; handed to sales as the highest-confidence pitch.

The hardest problems it surfaced weren’t Deal-Analyzer-specific, so I now drive them cross-functionally: deal-type templates, a generic Kanban, and email ingestion.

Step 7: learnings and next steps

What I’m carrying forward

AI should assist, not decide:
When the user is an expert, design to support the judgment, not make it. The counterintuitive move earned the trust.

AI should assist, not decide:
When the user is an expert, design to support the judgment, not make it. The counterintuitive move earned the trust.

In a data product, trust is the design: Provenance and honest partial states did more for credibility than any chart.

In a data product, trust is the design: Provenance and honest partial states did more for credibility than any chart.

Subtraction can provide clarity:
Eight tabs cut to three views; the scores taken off the page.

Subtraction can provide clarity:
Eight tabs cut to three views; the scores taken off the page.

Sequence the collaboration:
Align with engineering on what exists before high-fidelity; it changed the team's velocity, not just mine.

Sequence the collaboration:
Align with engineering on what exists before high-fidelity; it changed the team's velocity, not just mine.