Key Takeaways

  • Vertical document extractors are easier to sell early, but they trap engineering teams into maintaining narrow, brittle rule sets.
  • Human document reading depends on visual primitives like paragraph whitespace and table borders, which remain identical across industries.
  • Training models on extreme edge cases, such as distorted and faxed medical records, directly improves accuracy on clean financial and legal filings.
  • Reducto separates document processing into two distinct layers: zero-shot horizontal extraction for objective text, and promptable configurations for subjective business metrics.

The Trap of Vertical Extraction

Every founder building in document intelligence faces an early choice: target a single vertical like real estate leases or medical billing, or build a horizontal platform for all documents. Choosing a vertical feels safer. You get clear buyer personas, simpler scoping, and an immediate sales pitch. But as Reducto co-founder and CEO Adit points out, that early speed comes with a steep technical debt.

“It's a lot easier to build a vertical specific document processing platform,” (play from 14:24) Adit explains. “The trade-off that you make there is I think a lot of documents ultimately boil down to very similar primitives.” (play from 14:30)

When a team builds an extractor solely for mortgage forms or insurance claims, engineers end up hardcoding document-specific layout rules. The moment an insurance provider tweaks a column layout or shifts an approval block, the parser fails. Vertical solutions often confuse surface-level document templates with the mechanics of reading. Human beings do not learn to read contracts differently from how they read lab results. They learn basic visual grammar first.

Visual Primitives Are Universal

Document comprehension relies on visual layout rather than industry-specific vocabulary. A bank statement and a patient intake form look different, but human eyes interpret them using identical spatial rules.

As Adit explains, “You as a human when you look at a page make a lot of snap judgments. Like every time there's a gap between two paragraphs, you know, okay, this is a new semantic block of information.” (play from 14:42)

Whitespace signals a shift in topic. Margins create structure. Bounding lines organize data into rows and columns. These visual rules hold whether a page contains balance sheets or lab results. Instead of building narrow schema parsers, Reducto designed models like R1 to read these underlying visual building blocks. By approaching document parsing as a visual spatial task, the system extracts text and table structures without requiring custom training for every vertical.

Messy Faxes Make Clean Financials Better

The real test of a horizontal architecture comes from edge cases. If you train a layout model only on clean digital PDFs from investment banks, that model falls apart when an invoice contains a scan artifact or uneven tilt.

Reducto took the opposite path by exposing their models to some of the messiest pages in existence: faxed healthcare records. Faxes arrive distorted, skewed, and degraded by low-resolution noise. When a vision model learns to reconstruct a medication table through fax noise, reading a crisp corporate PDF becomes easy.

Adit points out this transfer effect directly: “The work that we do on making sure that we are resilience to messy tables and healthcare that were faxed means that we are naturally better at offering a reliable solution for our customers in finance as well.” (play from 15:08)

Training on the dirtiest inputs strengthens the baseline system. Solving the hardest visual hurdles in medical paperwork creates accuracy benefits across legal, finance, and logistics documents.

Separate Objective Pixels from Subjective Rules

The architectural problem is knowing where horizontal extraction stops and customer customization starts. Reducto draws a strict line between what appears on the page and how a specific team interprets that information.

Adit argues that extracting raw layout must remain horizontal and zero-shot: “We don't want to be in a situation where for objective tasks like parsing you can very clearly say on a given page like hey this is exactly what was on the page. We want that to be horizontal, we want that to be zero shot.” (play from 16:17)

If a page displays five rows in a table, the parsing engine must return those exact five rows without custom code. Customization belongs entirely in the business logic layered on top. “For things where it's a matter of opinion, like maybe a finance team has a different definition of a certain metric, yeah those things should be promptable and steerable and improve over time,” (play from 16:32) says Adit.

Keep visual extraction factual, universal, and zero-shot. Keep business definitions flexible and promptable.

What to Do With This

Audit your document pipeline this week to isolate layout parsing from business logic. Strip hardcoded field coordinates and bespoke regex from your ingestion layer, then run your core parser against five degraded, photocopied, or faxed documents to see whether your extraction relies on fragile assumptions.