writing suspicious activity reports (sars) is manual and repetitive.
analysts copy transaction facts into narrative templates, which is time-consuming and error-prone.
## what this project does
this prototype uses python templates to generate first-draft sar narratives from synthetic transaction data.
it then evaluates the drafts for completeness (who/what/when/why) and readability.
all data is synthetic and reproducible, created with faker.
## project structure
- data/ → synthetic csv + text templates
- docs/ → example sar drafts, evaluation results, plots
- notebooks/ → main jupyter notebook with workflow
- requirements.txt → dependencies (minimal for v1)
## how to run
1. create a virtual environment and install dependencies:
```bash
pip install -r requirements.txt
2. open the notebook:
jupyter notebook notebooks/sar_drafting_prototype.ipynb
3. run the cells to generate drafts, metrics, and plots.
results
dataset: 500 synthetic transactions
drafts: narratives generated from flagged alerts
evaluation: completeness checks + readability scores
visualization: histogram of readability distribution
next steps: adding spaCy for validation, rapidfuzz for novelty, and a variation layer for more natural drafts
disclaimer
this project uses only synthetic data.
it is for educational purposes only and is not a production compliance tool.