Managing your own money has become a lot harder since the number of digital payments in India has gone up a lot. The most popular tools still rely on fixed keyword heuristics, which means they can't learn, make predictions about the future, or give tax advice. FinMate fixes these problems with a cloud-native microservices platform that combines three machine learning models with a rule-based engine for optimising taxes. An XGBoost classifier puts incoming transactions into one of 18 spending categories and gets a weighted F1 score of 0.943, which is 32.8% better than a keyword baseline. A hybrid LSTM/ARIMA model makes spending predictions for 30, 60, and 90 days. It has a 30-day RMSE of ₹312, which is 23.4% less than a seasonal naive method.An Isolation Forest module spots strange transactions in real time with 91.3% accuracy, which is 14.3 percentage points better than a z-score detector. The tax engine connects different types of annual spending to deductions under the Indian Income Tax Act, including Sections 80C, 80D, HRA, 80E, and the standard deduction. This covers 85% of the most common types of deductions. An evaluation of 2,400 held-out transactions from a 12,000-record corpus, along with an ablation study and user acceptance testing (n=15), demonstrates a 73% reduction in time-to-financial-insight and a System Usability Scale score of 81.4. Anyone can see all the code, data, and a Docker deployment so they can make their own version.
