Lemon: fintech platform for SaaS financing
Role: Co-founder & Product Engineer
Focus: Product strategy, UX, frontend engineering (Next.js, React, Tailwind)
Outcome: Raised £500k, secured 4 lender partnerships (2 UK, 2 US), pivoted twice based on customer learning
Lemon started with a simple idea: make software buying easier for small businesses through "Buy Now Pay Later for SaaS". Imagine a "Pay with Lemon" button on checkout pages, inspired by Klarna.
The reality was messier and more interesting. Over two years, we went through three distinct product phases, each one teaching us something new about the problem we were trying to solve. I designed and built all three versions, working closely with our CTO on backend integrations whilst owning the entire frontend experience.
Phase 1: Subscription management app for SMBs
Users: ~20 small businesses using it regularly
Whilst we were negotiating lender partnerships (which took months), we knew we'd eventually need all three sides of the marketplace: SaaS vendors, lenders, and software buyers. We started with what we could ship immediately: the buyer side.
We launched a free SaaS subscription manager to help SMBs track their active subscriptions, licences, and payment methods. The thinking was that understanding their software spend would be valuable when we introduced financing options later.
We were right about the problem – most small businesses had little visibility into what they were spending on software. Around 20 businesses used it regularly.
The subscription page showing a table of SaaS subscriptions.
What I built
I started this before our CTO joined, so designed and built the MVP solo:
- Chose the tech stack (Next.js, Vercel, Supabase, Tailwind) which we used throughout Lemon
- Built user onboarding, team invitations, and subscription tracking
- Developed an AI feature allowing users to quickly track a subscription by uploading an invoice (used OpenAI API for document parsing)
- Launched to users and gathered feedback to shape the next phase
Design approach:
I worked directly in code using HTML & Tailwind – it's faster for me than mocking up in Figma first. For quick ideation, I'd either sketch rough wireframes in Google Slides (oddly effective for me) or use v0 ↗ to generate clickable prototypes. I'd typically work with ChatGPT first to refine the context and requirements, then feed a detailed prompt into v0 to get something I could iterate on or share with the team.
Individual subscriptions could be managed from a side panel. I designed and built the responsive UI.
Phase 2: Financing platform for SaaS vendors
Users: Multiple UK SaaS vendors onboarded as early adopters
As lender integrations came online, we realised there was stronger pull from sales-led SaaS vendors. They didn't need a checkout widget – they wanted to offer flexible payment terms in their outbound deals.
We built a vendor-facing financing platform where sales reps could create a deal for a customer and have Lemon handle lender matching, eligibility checks, and repayment calculations.
How it worked
- Sales rep signs in and creates a new company
- Lemon auto-fetches company details and financials via Companies House API
- The app generates an indicative financing offer (or declines if unqualified)
- Reps use a calculator UI I designed to choose whether to absorb or pass on the financing fee
- Lemon generates a unique application URL for the buyer
- Buyer completes the application and is sent to the lender via their API
- Vendor tracks deal progress and buyer updates through Lemon
The financing fee decision was product-critical: If a customer needed monthly payment terms to close a deal, the sales team could choose to swallow the lending fee. But if customers were asking to pay monthly instead of annually (a preference, not a necessity), they'd typically be willing to pay the fee to get preferred payment terms.
We onboarded several UK SaaS vendors as early adopters, but final lending decisions sat entirely with the partner banks, which made deal conversion unpredictable. That dependency exposed a flaw in the model and pushed us to rethink who we were really building for.
What I built
Frontend architecture:
- Designed and built the frontend in a Next.js monorepo with Tailwind CSS
- Created a shared Tailwind UI component library (based on Tailwind Plus's Catalyst components, which accelerated building new features
- Added Storybook for component documentation and visual testing
- Example customisation: extended the base input component to support currency prefixes, which we needed throughout the app
Design decisions:
- Used side panels for deal details to avoid page refreshes and maintain context
- Drew inspiration from Linear (for handling complexity cleanly) and Attio (for data-heavy table interfaces)
- Focused on keeping the UI uncluttered despite the complexity
Collaboration:
- Worked closely with CTO on lender integrations and business logic
- Designed, wrote copy for, and built the marketing website
The UI that helps decide who would pay the lending fee.
Phase 3: AI-powered platform for finance brokers
Outcome: ~50 brokers on waiting list; HubSpot app integration built
Whilst building and iterating the SaaS financing platform, we were still doing a fair amount of manual work (do things that don't scale ↗). We were effectively acting as finance brokers ourselves – collecting data, chasing documents, liaising with lenders.
Doing this unscalable work revealed where the real pain lived: inside the broker workflow.
Understanding the problem
Initial calls with brokers showed us that brokering is relationship-heavy and admin-heavy in equal measure. One senior broker told us he'd spend most evenings on the sofa with his laptop, updating systems with the day's conversations and progress.
There was a clear problem around managing the admin side of brokering – work they didn't enjoy. Our hypothesis: an AI-powered app could automate some (and eventually all) of the admin work.
The vision: A system that could intake new financing deals via natural language, perform research on the company to help validate them, create a data room, autonomously collect and verify data/docs from the client, then make all the research and docs available to the broker for easy analysis and lender application.
A critical pivot
Through further customer research and sales calls, we quickly realised brokers did not want to manage yet another software app for their process. Most brokers in our ICP already used HubSpot, so we made the decision to build a HubSpot app that surfaced Lemon's capabilities directly where they were already working.
We focused Lemon on being an AI-first platform for finance brokers that automated repetitive admin and data validation.
Key features we built
AI Data Rooms:
Brokers could define which documents they needed from clients. Lemon created a client-facing upload portal, and LLMs verified each submission (e.g., detecting if only 12 months of bank statements were provided when 24 were required).
Companies House integration:
Instant business verification and financial overview to accelerate due diligence.
AI research agent:
Summarised information about the client company based on autonomous web research, saving brokers time and highlighting potential risk signals.
HubSpot app:
Instead of replacing their CRM, we built a HubSpot app that surfaced the same AI insights directly within the company view where brokers already worked.
The HubSpot app allowed brokers to view Lemon-derived data directly in their company page, meeting them where they already worked
What I built
Product and frontend:
- Designed and built both the broker web app frontend and the HubSpot app UI
- Reused the component library from the financing app to move fast and maintain design consistency
- Integrated AI models for document checking and company research (OpenAI APIs, prompt engineering, cost optimisation)
- Used Langfuse to monitor model behaviour and improve accuracy
- Worked closely with CTO on backend data flows and APIs
Design process:
- Prototyped quickly using Cursor and ChatGPT, shared early for feedback
- Focused on presenting dense data clearly without overwhelming the user
- Drew continued inspiration from Linear and Attio for handling complex workflows elegantly
Although the broker app never reached live users (we wound down Lemon before launch), we had around 50 brokers on a waiting list ready to join.
Technical highlights
Component library & design system:
- Built a shared UI component library based on Tailwind Plus's Catalyst components, customising to fit our needs
- Used Storybook for documentation and visual regression testing
- Enabled consistent, fast feature development across multiple product pivots
AI integration:
- Built features using Vercel's AI SDK (document verification, company research, invoice parsing), which kept us LLM-provider agnostic
- Learned prompt engineering through iteration – testing different approaches to improve accuracy
- Implemented Langfuse to monitor model behaviour and understand where prompts were failing
- Saw promising early results: document verification caught missing bank statements, company research surfaced useful risk signals
Full-stack pragmatism:
- Whilst I owned the frontend, I worked across the stack when needed
- Set up database schemas, integrated third-party APIs (lenders, Companies House, HubSpot)
- Deployed and monitored production apps on Vercel
What I learned
Pivot fast, but pivot with evidence:
Each phase change was driven by real customer conversations and usage data, not gut feel. The subscription manager taught us about SMB pain points. The vendor platform revealed the dependency problem with lenders. The manual broker work showed us where the real value was.
Build what you can ship, not what you wish you could:
Starting with the subscription manager whilst lender partnerships were being finalised meant we were learning and building an audience from day one, not waiting months to start.
Meet users where they are:
The HubSpot integration was the right call – brokers didn't want another tool, they wanted their existing tools to work better. That insight only came from talking to users early and often.