The Problem
Most personal finance apps do the same thing. they connect your bank and show you a coloured pie chart of your spending. The problem is nobody looks at the pie chart. Categorisation is inaccurate, the charts don't answer the question "can I afford X?", and there's no feedback loop to change behaviour.
We wanted to build something fundamentally different: a finance app that talks to you, understands your spending context, and gives you honest, personalised answers to real financial questions.
What We Built
Bank feed aggregation via Plaid. users connect their UK and EU bank accounts through Open Banking. Transactions are pulled in real time (webhooks) and stored in a normalised transaction ledger.
AI transaction categorisation: rather than regex-based categorisation (which misclassifies ~25% of transactions), we use GPT-4o with a structured output schema to categorise transactions. The model sees the merchant name, amount, timestamp, and recent transaction history for context, and returns a category, subcategory, and confidence score. Low-confidence transactions are shown to the user for one-tap correction, which feeds back into a fine-tuning dataset.
Spending intelligence: the app computes a weekly "financial health score" across five dimensions: spending vs income ratio, emergency fund progress, debt trajectory, recurring commitment growth, and discretionary control. Each dimension has a 0-100 score and a plain-English explanation. The score updates nightly after new transactions sync.
Conversational budget coach: the centrepiece feature. Users ask anything: "Can I afford a £1,200 holiday in July?", "How much do I spend on food per month vs last year?", "What would happen to my savings if I cut my coffee spending in half?" The AI coach runs queries against the user's actual transaction data and answers in conversational language with concrete numbers.
Savings goals: users create goals (house deposit, emergency fund, holiday) with a target amount and timeline. The app shows current trajectory and the required monthly saving to hit the goal, updated weekly.
My Role
Everything technical:
- Plaid integration and webhook processing pipeline
- AI categorisation service (FastAPI + async processing queue)
- Financial computation engine (spending patterns, health score, goal projections)
- Conversational coach with RAG over user's transaction history
- Next.js web app and React Native app (shared API, diverging UIs)
- Supabase database with RLS, column-level encryption for PII
- Product decisions: led user research interviews, prioritised feature roadmap
Outcome
Closed beta with 180 users over 4 months. Users tracked £280,000+ in savings goals collectively. Retention at 30 days was 61%. unusually high for a finance app, which we attribute to the conversational coach feature driving daily engagement.
Average user accessed the budget coach 4.2 times per week. Most common query type: "Can I afford..." (42%), followed by spending comparisons (28%) and savings projections (30%).
Currently raising a pre-seed round.