DeepHighlight

What if we could target key parts directly from the document source? No insight left behind. Today’s deep research agents read text—but documents are more than words.

Comment

Inspiration

Most current AI tools are designed for text-based conversation only—you ask a question, the AI responds with an answer, and that’s it. But when it comes to complex documents like mortgage contracts, insurance policies, or legal agreements, that’s simply not enough.

These documents include not just words, but figures, tables, formatting cues, and visual context. Professionals don’t just “read” them—they analyze, highlight, and annotate them visually. Yet today's AI agents:

  • Can’t highlight the exact sentence or clause they’re referring to
  • Can’t capture nearby figures, tables, or labels that are essential to understanding
  • Can’t create visual trust through contextual evidence

This is the key insight behind DeepHighlight:

“An image says more than a thousand words—and a highlight says more than a thousand tokens.”

We built DeepHighlight to close that gap. It doesn’t just give you an answer — it shows you where the answer lives in the document, highlights the surrounding context, and lets you add memos just like a human reviewer would, just like how professionals manually highlight and annotate contracts.

What it does

DeepHighlight allows users to:

  • Ask questions about the content using template system instantly.
  • Automatically receive answers along with highlighted evidence from the document.
  • Download the annotated PDF with all highlights included.

For example, a bank customer service always got some question that related to mortgage promotions. Nowadays people spend time to read huge number of pages. With the app, they could add FAQ to template system, then instantly ask which discount or promotional offer will be applied, and instantly see not only the answer but where it is stated in the document as a reference.

How we built it

  • Initial the project Kiro spec driven AI development
  • Continues to add tasks inside spec
  • Vibe coding

Challenges we ran into

  • Ensuring the highlights don't overlap or clutter the page took multiple iterations of design and testing.
  • Balancing AI accuracy with speed, especially when users ask complex or ambiguous questions, was tricky.

What we learned

  • Users don't just want answers—they want proof and traceability in the original document.
  • Visual feedback like highlights and memos dramatically increases trust in AI responses.
  • There is a strong need for tools that combine AI reasoning with document interaction, especially in regulated industries.

Built With

Share this project:

Updates