BankFlow AI

Problem

Raw Open Banking transaction feeds are noisy: the merchant names are truncated, the categories are wrong, and the patterns that matter (unusual spend, recurring charges, cash flow problems) are buried in the noise. Finance tools exist, but they don't apply LLMs to make sense of the data in plain language.

Approach

Connect to TrueLayer's Open Banking API to pull normalised transaction data. Apply LLM-based semantic categorisation (replacing brittle rule-based matching), run anomaly detection over rolling windows, and surface the results in a clean dashboard.

Key technical decisions

TrueLayer chosen for its UK Open Banking coverage and developer-friendly sandbox. LLM categorisation preferred over ML classifiers because it handles novel merchants without retraining and produces human-readable explanations for each categorisation.

Outcome

In progress. Would be measured by categorisation accuracy vs. manual labels and anomaly precision/recall on synthetic test transactions.