Inspiration
Millions of people in the UK are paid irregularly: delivery riders, gig workers and people on zero-hours contracts. Many of them are also new to the country and manage money in a second language. Budgeting apps tell them where their money went, but the hard part is knowing where it's going: the week when rent, an annual bill and a slow pay week land together.
We built Future You for people like Adam, a delivery rider in Manchester whose first language is Arabic. Instead of a red warning, he gets a message from his future self, in his own language and his own voice, telling him what's coming and one small thing he can do now.
What it does
- Forecasts 90 days ahead from past transactions, showing a typical-weeks line and a slow-weeks line.
- Spots risk weeks and explains why: "Car insurance and rent are due in the same week."
- Plays a message from Future You in Arabic or English, with English subtitles.
- An AI coach you can type or talk to, which only suggests options that have already been calculated, and points to free help (StepChange, Citizens Advice, MoneyHelper).
- "Simulate next week" and "Can I afford it?" to explore what-ifs.
How we built it
- Frontend: Lovable (TanStack Start, TypeScript), with right-to-left Arabic support.
- Backend: n8n Cloud webhooks:
/forecast,/coach,/future-message,/voice-speak,/voice-transcribe. - AI: Claude, with a fast model for coach replies and a stronger one for Future You scripts and translation.
- Voice: ElevenLabs text-to-speech (
eleven_v3) and speech-to-text (scribe_v2), with Arabic and English voices.
The forecast is maths, not AI
The forecast is plain JavaScript with no AI, so it gives the same answer every time. From the weekly pay history $p_1, \dots, p_n$ we take a slow week and a typical week:
$$ I_{\text{low}} = P_{25}(p), \qquad I_{\text{typical}} = P_{50}(p) $$
Then we walk forward day by day. For the slow-weeks line:
$$ B_{t+1} = B_t - \bar{s} + I_{\text{low}}\,\mathbb{1}[\text{payday}] - \sum \text{bills}_t $$
where $\bar{s}$ is average daily everyday spending. If $\min_t B_t < 0$, that period is a risk. The shortfall and fix are:
$$ \text{shortfall} = \left\lceil -\min_t B_t \right\rceil, \qquad \text{save per week} = \left\lceil \frac{\text{shortfall}}{\text{weeks until}} \right\rceil $$
For Adam that's $\lceil 143 / 9 \rceil = £16$ a week for 9 weeks.
Keeping the AI honest
The AI never calculates numbers; it only explains them. The code works out every figure and every option's effect, then gives Claude a FACTS object. A guard node checks every £ amount in the reply. If an amount isn't in FACTS, the reply is rejected and a safe fallback is used. A second guard checks that the Arabic translation kept every amount. Every message is labelled as AI and as not financial advice.
Challenges we ran into
- AI inventing numbers. Prompting alone wasn't enough, so we added code-level guards and approved cached scripts as a fallback.
- Language drift. An English question sometimes got an Arabic answer, and "next week" was used for something two months away. We fixed both with explicit language and date context.
- Two sources of truth. The app and backend gave different shortfalls (£127 vs £143) because they used different "today" dates. We made the coach use the forecast the app sends.
- Model quirks. One model spent its whole token budget thinking and returned no text, so we had to tune the API settings for each model.
- Integration across four people. Keeping the frontend, backend, forecast and voice on one agreed API contract, plus CORS, syncing Lovable with GitHub, and shared access between accounts.
What we learned
- Deterministic code plus an LLM for explanation is much more trustworthy than an LLM alone for money.
- Guardrails belong in code, not just in prompts.
- Agree the data contracts first: it made four people's work fit together.
- Inclusion goes beyond translation: plain language, a familiar voice, no judgement about money sent home, and a clear route to free help.
What's next
- Proactive alerts when a new risk appears
- More languages and personas
- Consent-based voice cloning, so Future You sounds like you
- Running locally with an on-device LLM and voice model, for privacy
Built With
- claude
- elevenlabs
- github
- javascript
- json
- lovable
- n8n
- react
- tailwind
- tanstack
- typescript
- webhooks