Welcome to RemediationGPT

To showcase how GenAI can be used to make data from large projects easily searchable, we've built RemediationGPT on publicly available reports from the Impartial Reviewers’ analysis of a remediation project that has been ongoing at a large Danish bank over the past years.

Reports 1, 2, 5 and 6 are available in fully digital form. Furthermore, reports 1, 2 and 5 are already available in English, whereas reports 3, 4 and 6 are only available in Danish. Additionally, reports 3 and 4 are scanned PDFs.
We've therefore had to apply additional steps when preparing the data from 3, 4 and 6, and the quality of answers based on those reports is expected to be lower than when referencing reports 1, 2 and 5.

If the confidence in retrieved report sections is too low, the assistant will not return an answer. In such cases, try rephrasing your question.

Ask about debt collection remediation reports.
Example: "What is Issue 26?" or "How is tax compensation handled?"
Note that it may take around a minute for the GPT to answer when you first ask

... also, it only knows what's in the reports, so don't waste your time and our OpenAI tokens asking about the meaning of life!

This chatbot was built through a multi-stage pipeline:

  1. PDF to Markdown: Converted scanned and digital PDFs into clean text.
  2. Manual Cleanup: Fixed layout glitches and OCR/translation errors.
  3. Smart Chunking: Split documents by deep section headers (e.g. 9.4.2.5.1.1.6).
  4. Metadata Enrichment: Extracted keywords, legal references, issue numbers, definitions, and cross-references.
  5. Postprocessing: Validated structure, trimmed large chunks, reran metadata if needed.
  6. Embedding: Embedded chunks using OpenAI + FAISS for semantic retrieval.
  7. Deployment: Hosted backend on Render and frontend on GitHub Pages.