For a UK consumer-fintech startup, I built the AI assistant inside their money app: seven specialised agents orchestrated in LangGraph, every financial figure computed by deterministic code rather than the model, and a four-layer compliance architecture designed to keep the assistant on the information side of the FCA advice boundary — tested against 100+ real user survey queries.
| Value | What it is |
|---|---|
| 43 | Deterministic tools, 0 hallucinated figures by design |
| 7 | Specialised agents |
| <100ms | Streaming first token |
| 100+ | Real-survey compliance suite |
A UK consumer-fintech startup wanted an AI assistant inside their money app — connected to users' real bank transactions through open banking, able to answer "what did I spend on groceries last month?", surface forgotten subscriptions, and point out cashback left on the table.
Two failure modes kill a product like this. The first is hallucination: ask a bare LLM about last month's spending and it will happily estimate. Get a number wrong once about someone's own money and the trust never comes back. The second is regulatory: under UK FCA rules, an app that drifts from financial information into financial advice without authorisation has a legal problem, not a product bug.
Over five months I designed and built the assistant's engine — the multi-agent orchestration, the financial tool layer, and the compliance architecture — on LangGraph and LangChain, integrated with TrueLayer for open-banking data and Tillo for gift-card offers.
The orchestration is an eleven-node LangGraph graph running seven specialised agents — budget, savings, wellness check-in, onboarding, general, gift-card recommendation, and the classifier that routes between them. Routing is conditional and self-correcting: an agent handed a query outside its lane doesn't guess. It flags the query as out of scope and bounces it back to the classifier, which reroutes to the right agent — so the product feels like one assistant, not six bots taking turns.
| Step | What happens |
|---|---|
| Guardrails in | Prompt-injection detection and PII redaction, self-hosted local models |
| Classifier routes | Intent from the full conversation history, to one of seven agents |
| Tools compute | 43 deterministic pandas tools produce every financial figure |
| Compliance out | Two-step advice check, then the response streams over SSE |
The core rule of the tool layer: the LLM never computes a financial figure. All arithmetic runs through 43 deterministic pandas tools — spending summaries, merchant analysis with fuzzy matching (so Tesco Extra, Tesco Metro and TESCO SUPERMKT read as one merchant), anomaly detection, and subscription detection that reframes cost annually, because £15.99 a month feels small and £191.88 a year doesn't. The model chooses the tools and writes the sentences; the numbers are exactly right by construction, not plausibly close.
"A SIPP is a pension wrapper" is information; "open one" is advice — the FCA boundary is semantic, and keyword filters can't see it. A two-step check classifies every response against a 341-line compliance prompt of worked examples; anything classed as advice becomes a mandated deflection script. Better to wrongly block than wrongly advise.
Every calculation is auditable: same input, same output. Every figure a user sees comes from a deterministic pandas tool over their real transactions. Where most assistants let the model "reason about" numbers, this one structurally cannot invent them.
Ask the budget agent for a joke and it doesn't improvise — it signals the query as out of scope and hands it back to the classifier, which reroutes. Agents stay specialised without the harsh "I can't help with that" dead ends that make multi-agent products feel stitched together.
Prompt-injection detection and PII redaction run as self-hosted local models — no third-party API calls, so sensitive data never leaves the platform and the checks add no network latency. The full advice check costs ~500ms, and SSE streaming puts the first token in front of the user in under 100ms.
| Metric | Outcome |
|---|---|
| Hallucinated figures | None by design — every figure computed by deterministic code, not the LLM |
| Advice boundary | Designed to keep the assistant on the information side of the FCA advice boundary — LLM-as-judge suite over 100+ real user survey queries |
| Latency | <100ms first token, ~500ms compliance check (pre-deployment testing) |
| Scope | 7 agents, 11 LangGraph nodes, 43 tools; TrueLayer and Tillo integrations; built over five months |
| Status | Delivered ahead of the app’s public launch — the figures above are from pre-deployment testing |
This work sits under NDA, so the client and app are anonymised. And the compliance layer is an engineering claim, not a legal one: designed and tested to keep responses on the information side of the FCA boundary — which is not regulatory sign-off, and no substitute for FCA authorisation where regulated advice is the goal.
A designed PDF version of this case study is in this repo: 09-fintech-assistant.pdf.
Freelance AI engineer — Expert-Vetted on Upwork (top 1%), 100% Job Success over 70+ projects, $400K+ earned, 5,750+ hours billed. I build production LLM systems for regulated industries: insurance, finance, law.