What this video shows
Linkov describes claim material as handwriting, faxes, duplicate pages, and PDFs that can reach gigabytes. He walks the process of a claim and argues that cleaning this data is both the largest barrier and the largest opportunity. He stages the work from data cleaning, to analysis inside a single document, to insights that compare many claims.
Wisedocs built a benchmark with Artificial Analysis to test how models handle medical records, which gives the talk a measured component beyond a single demo. The speaker sells a document-review product, so the architecture reflects that product. The staged approach is useful for any messy-document pipeline, but the quality numbers come from the vendor's own benchmark and should be reproduced on your documents.
Read Wisedocs for the document-review product the talk is built around. Check Artificial Analysis, the analysis provider that co-built the medical-records benchmark.
What you will learn
- Treat data cleaning as a first-class stage because the model cannot recover from garbled input.
- Separate within-document analysis from cross-document analysis because they need different checks.
- Track which source formats produce the most errors so cleaning effort goes where it pays.
- Reproduce a vendor benchmark on your own documents instead of trusting its headline numbers.
How to apply this safely
- Sample real documents and record their formats, sizes, and defect types.
- Build a cleaning step and measure how often it produces readable, correctly ordered text.
- Test within-document extraction against known answers, then test cross-document questions separately.
- Store the source hash with each extracted record so a wrong answer can be traced back to the document.
Important limitations
- The speaker works for Wisedocs and presents its pipeline and benchmark, so the results describe that product.
- The medical-records benchmark was built in partnership with an analysis firm and may not generalize to other document types.
- Insurance rules differ by jurisdiction, so accuracy targets and retention requirements must come from the operator, not the video.
Sources to check
- Wisedocs The document-review platform presented in the talk.
- Artificial Analysis The analysis provider that partnered on the medical-records benchmark.
Continue learning on Learnetto
AI evals guide
Build datasets and graders for extraction and document tasks.
Long context or retrieval
Decide whether to retrieve passages or send whole documents.
Best AI engineering courses
Study data, evaluation, and production trade-offs together.