Bank Statement Parser Agent

Overview

This Python agent automatically generates a custom parser for any bank statement PDF. It uses a self-correcting loop to generate Python code, test it against sample CSV data, and refine the parser up to 3 times. The generated parser function, parse_pdf(pdf_path: str) -> pd.DataFrame, returns a pandas DataFrame that exactly matches the expected CSV schema. This allows evaluators to run the agent on new bank statements without manual code changes.

Agent Architecture

The agent operates in a self-correction loop. It begins by reading a sample PDF to understand its text structure and a target CSV file to understand the desired output format. It then constructs a detailed prompt, including this context, and instructs the Gemini model to generate a Python parser. Immediately after creation, the agent saves the parser and executes a test to validate that the parser's output structure (e.g., dictionary keys, return types) matches the target CSV schema. If the test fails, the agent captures the error output, adds it as feedback to the prompt, and asks the model to generate a corrected version. This cycle repeats up to three times until the test passes or the attempts are exhausted.

Agent Diagram (Loop Description)

┌───────────┐
│  PLAN     │
│ Generate  │
│  LLM      │
└─────┬─────┘
      │
      ▼
┌───────────┐
│ GENERATE  │
│  Parser   │
└─────┬─────┘
      │
      ▼
┌───────────┐
│  TEST     │
│ Compare   │
│ DataFrame │
└─────┬─────┘
      │ Fail? ── Yes ──▶ Feedback → PLAN
      │
      ▼
     Done

---

Setup Instructions

1. Clone Repository

git clone https://github.com/apurv-korefi/ai-agent-challenge.git
cd ai-agent-challenge

2. Install Dependencies

pip install -r requirements.txt
# If requirements.txt not provided:
pip install pandas pdfplumber google-generativeai

3. Set API Key

export GEMINI_API_KEY="your_gemini_api_key_here"

(Windows CMD: set GEMINI_API_KEY=your_key)

4. Run Agent

Generate a parser for a target bank (e.g., ICICI):

python agent.py --target icici

This will:

• Read data/icici/icici_sample.pdf and data/icici/icici_sample.csv
• Generate custom_parsers/icici_parser.py
• Run automated contract tests and self-correct up to 3 times if needed
• Log results to logs/icici_parser_generation.log

5. Test New Bank

To generate a parser for a new bank (e.g., SBI):

1. Place a sample PDF and CSV in data/sbi/sbi_sample.pdf and data/sbi/sbi_sample.csv.
2. Run:

python agent.py --target sbi

The agent will automatically generate custom_parsers/sbi_parser.py and test it.

---

Notes

• The agent ensures the generated parser strictly adheres to the CSV schema.
• Logs in logs/ capture every attempt, including the generated code and test results.
• The parser uses pdfplumber and regex for robust extraction of dates, amounts, and descriptions.

---