Korva is a scroll of the internet’s best ideas.
Messy files
Group underwriting starts with a broker asking several insurers for bids and sending an email with PDFs and spreadsheets: who the employees are, what claims they have, and what the current insurance looks like. Every one comes in a different spreadsheet format or PDF layout. Software with universal importers and universal intake kind of worked.
What works amazingly is giving the files to an agent and having it write Python to get them into a standard format. The agent tests the script and loops until it works, then the script is discarded. Each file gets freshly generated, single-use code that never runs again.
Beyond spreadsheets
Brokers, providers, and employers used to be required to use Curative's standard format, and people in provider offices reformatted spreadsheets by hand. Now they can send whatever they have, even scribbles on a napkin, and the model converts it in about 15 minutes. Data ingestion pipelines that once took hundreds of people moving spreadsheets around are now a model writing Python.
The same trick works elsewhere. Models are terrible at editing Word documents directly, but when told to write Python to edit one, they do it well. They can also generate CAD models by writing Python that draws the part, even though they are not good at 3D visualization or outputting drawings as vectors.
Why is throwaway code better than universal importers?
A universal importer has to anticipate every layout it might meet, so it breaks on the first format nobody planned for. Generated code is written for one specific file and checked against it. Because producing code costs very little with a model, the script need not be reusable. The cost of one-off logic is small compared with the cost of maintaining a general parser.
How do agents test their own scripts?
An agent runs the script on the file it was written for and reads the output or error message. It compares the result to the target format, then revises the code and runs it again. This loop repeats until checks pass, such as matching column counts or valid values. Software developers call the same practice test-driven iteration.
How can a model make CAD parts by writing code?
Some CAD tools define parts through code instead of mouse clicks. OpenSCAD, released in 2010, and CadQuery are examples, where a script lists shapes, dimensions, and operations such as cuts and unions. A model that writes that script is describing the part in text, which it handles well. Running the script produces the 3D geometry.
More from 20VC with Harry Stebbings
