AI-Integrated Cash Flow ETL Pipeline
Engineered an ETL pipeline utilizing n8n and a local Ollama LLM to extract, transform, and structure raw financial data for analytics.
The Objective
Manual financial data entry creates significant friction for small businesses. Designed as a scalable solution for the UMKM sector, this system automates the ingestion and categorization of unstructured transaction records while maintaining strict data privacy through localized AI processing. To validate its real-world viability, the proof-of-concept was field-tested and deployed in collaboration with UMKM 32D Studio.
Data & AI Stack
- • n8n (Data Orchestration / ETL)
- • Ollama / Llama 3.1 (NLP Extraction)
- • Python / Pandas (Model Evaluation)
- • Supabase (PostgreSQL Storage)
System Interface
Frontend interface deployed for user input and analytics visualization.
Data Pipeline Architecture
ETL flow mapping: Raw NLP ingestion (Extract) → LLM categorization (Transform) → Relational database storage (Load).
Model Evaluation & Metrics
To rigorously evaluate the LLM's viability for automated data entry, a Python evaluation script was engineered using Pandas. The AI outputs were mathematically validated against a 100-row Ground Truth dataset. While the model achieved a 0% input failure rate and strong accuracy in extracting transaction types (88.1%) and exact nominal amounts (83.1% with a strict ± Rp 1 tolerance), the categorical classification (71.2%) revealed edge-case vulnerabilities in specific transaction scenarios, providing a clear roadmap for future prompt engineering.