Dynamic Reader is a working experiment in a simple question: can we make document mapping as intuitive as writing a prompt? Instead of training a classifier over labelled examples, it treats extraction and mapping as an LLM optimization problem — letting a DSPy pipeline learn the mapping itself, in a controlled, low-code environment. The goal is to move teams in text-centric fields (law, tax, compliance) away from black-box classification and toward tools where they directly manipulate the prompts that drive the mapping.
The shift is subtle but real. Traditional ML treats the problem as a numbers game of optimization; embedding LLMs with more data makes them inherently better at mapping, so the bottleneck moves from feature engineering to prompt design. Dynamic Reader is built as a demo of that direction — structured extraction and document intelligence you can inspect, steer, and reuse.
Check out the demo here:
