# "Gemma 4 Analyzed My Bank Statements – Apparently I 'Have a Problem' with Coffee and Late-Night Apps" > Source: > Published: 2026-05-22 12:27:10+00:00 This is a submission for the Gemma 4 Challenge: Build with Gemma 4 What I Built Bank statement Analyzer — upload 3–6 months of statements, get a breakdown of spending patterns, subscriptions you forgot about, anomalies, and concrete suggestions to cut your costs. Demo Code AbelCodeCanvas / my-bank-app Bank statement Analyzer — upload 3–6 months of statements, get a breakdown of spending patterns, subscriptions you forgot about, anomalies, and concrete suggestions to cut your costs. markdown 💰 Bank Statement Analyser Upload 3–6 months of bank statements and get a clear breakdown of: - 📊 Spending patterns – where your money really goes - 🔁 Subscriptions you forgot about – recurring charges you might not need - ⚠️ Anomalies – unusual or unexpected transactions - ✂️ Concrete suggestions – actionable advice to cut costs Powered by Gemma 4 26B A4B instruction‑tuned model via Hugging Face. 📋 Prerequisites Before you begin, make sure your local machine has: - Python 3.9 or higher (recommended: 3.10) - Git – to clone the repository - A Hugging Face account (free) with a User Access Token Create one here - At least 16 GB RAM (32 GB recommended) - GPU with 12+ GB VRAM (optional but strongly recommended for fast inference) – if no GPU, the app will fall back to CPU (very slow for 26B model) Note: The 26B A4B model is large but uses Mixture‑of‑Experts to reduce compute… How I Used Gemma 4 For my Bank Statement Analyser, I used Gemma 4 26B A4B (the instruction-tuned variant) on Hugging Face. While not exactly one of the standard sizes (E2B, E4B, or 31B Dense), this 26B parameter model strikes an ideal balance for the task: Long context handling – Bank statements over 3–6 months contain hundreds of transactions. The model’s large context window lets me feed entire statements without chunking, preserving temporal patterns. Structured extraction – Gemma 4’s instruction-tuning excels at parsing semi-structured data (PDF/CSV statements) and outputting consistent JSON breakdowns of spending, subscriptions, and anomalies. Reasoning for suggestions – The 26B size provides enough reasoning capacity to identify cost-cutting opportunities (e.g., duplicate subscriptions, high-fee accounts, irregular charges) without the latency or cost of a dense 31B model. A4B efficiency – The Mixture-of-Experts (A4B) architecture reduces compute per token, making it feasible to run locally or on a free Hugging Face T4 GPU. In short, Gemma 4 powers the entire pipeline: statement parsing → spending categorization → anomaly detection → actionable recommendations.