Two reference corpora described the same domain under incompatible specifications. No crosswalk existed, and the attribute scales were too dissimilar for hand-written rules. I extracted both from scanned PDFs, loaded them into PostgreSQL, and normalized the data. I then fit a model based on percentile mapping: I built empirical distributions for every numeric attribute on both sides, and an algorithm to map records from the legacy corpus to the modern one. Conversion encodes a source record's parameters as a percentile vector, then decodes it against the target distribution. Descriptive text went through constrained LLM prompting and was parsed back to JSON. The deliverable is a CLI client with lookup, search, and several render formats.
Built with
- Python
- PostgreSQL
- Statistical Modeling
- Prompt Engineering