CHAPTERS
- 0:00 – 0:42
Reducto’s mission: turning messy enterprise documents into LLM-ready structured data
The founders explain that Reducto converts complex documents—like insurance claims, health records, and financial statements—into clean structured data. The primary goal is to make downstream LLM workflows (RAG, summarization, reasoning) more reliable by fixing ingestion quality at the source.
- •Transforms complicated documents into structured outputs
- •Optimized for LLM use cases like RAG and summarization
- •Targets high-stakes enterprise document types (claims, records, statements)
- 0:42 – 1:30
Early traction with major customers and why ingestion is a core AI bottleneck
Reducto quickly gained adoption among notable companies (e.g., Vanta) and large, often unnamed enterprises. The founders frame document ingestion as a fundamental bottleneck for building effective AI applications, even for teams with strong engineering capabilities.
- •Rapid enterprise adoption within ~1 year of founding
- •Works with both startups and very large enterprises
- •Ingestion quality is a foundational constraint for AI apps
- •Many teams underestimate the time and complexity involved
- 1:30 – 2:57
From LLM memory tooling to document ingestion: the pivot that wasn’t planned
They started with a different YC application idea—long-term memory for LLMs—then discovered customers also needed uploaded files managed. A quick, rough segmentation demo unexpectedly drew strong demand, pulling them into document processing as a product.
- •Original product: long-term memory for LLMs
- •User requests expanded to file/document handling
- •A weekend segmentation prototype triggered strong inbound interest
- •Customer pull revealed how widespread the ingestion pain was
- 2:57 – 4:43
Why PDF processing stays unsolved: “schlep blindness” and the ‘ingestion team’ framing
They discuss how PDF extraction is widely known to be painful, yet few teams want to tackle it—making it an unglamorous but valuable wedge. Reducto positions itself as the dedicated ingestion team so application builders can focus on their core product.
- •Teams don’t want to become “PDF processing companies”
- •Schlep blindness: everyone knows it’s painful, nobody wants to fix it
- •Reducto abstracts ingestion as a service/API for product teams
- •Customer value: fewer internal resources wasted on parsing and cleanup
- 4:43 – 6:22
A vision-first approach: treating document parsing as a computer vision problem
Instead of relying on brittle rules and PDF standards, Reducto focuses on understanding documents the way humans do. The founders explain how layout cues (spacing, indentation, hierarchy) carry meaning and require robust modeling across a long tail of formats.
- •Replaces rule-based parsing with computer-vision-driven understanding
- •Uses visual cues like gaps, tabs, and hierarchy to infer semantics
- •Designed to work across many document types, not just invoices
- •Long-tail generalization is difficult but central to the product
- 6:22 – 7:49
Why better ingestion boosts LLM outcomes: accuracy gains and newly possible features
Customers often see substantial improvements in end-to-end LLM accuracy simply by switching ingestion providers. Reducto also reduces the need for heavy post-processing and enables use cases that previously failed due to scanned docs, missing metadata, or performance constraints.
- •Swapping ingestion can yield large LLM accuracy gains (reported up to ~30%+)
- •Improved outputs reduce downstream cleanup and chunking work
- •Enables processing of scanned documents and other hard inputs
- •Unlocks features that were previously impossible due to poor extraction
- 7:49 – 9:17
Handling the hardest real-world documents: handwriting, checkboxes, and visual annotations
They describe examples like extracting handwritten, messy real-estate questionnaires with very high accuracy. Reducto combines deterministic layout methods with modern VLM capabilities for the long tail, including cases like highlighted/circled table values and instruction-guided tweaks.
- •Handwriting extraction improved by modern VLMs vs traditional OCR
- •Supports heterogeneous inputs: checkboxes, handwriting, scans
- •Can handle visual edge cases (e.g., highlighted/circled numbers)
- •Customers can guide behavior with plain-text instructions
- 9:17 – 11:24
Closing a Fortune 25 customer fast: from Launch YC demo to enterprise gauntlet
The founders recount an intense sales cycle that began during the YC batch after a successful Launch YC demo. Despite not being salespeople, they navigated a months-long evaluation involving many stakeholders and direct comparison against the customer’s internal doc-processing team.
- •Deal originated from Launch YC and a failing document in the customer pipeline
- •Focused on being “substantially better” via a strong layout model
- •Competed primarily against an internal document processing team
- •Enterprise evaluation included extensive in-person grilling and multiple stakeholders
- 11:24 – 12:45
Reaching SOTA: building high-quality datasets and an internal data engine
They explain that progress required better benchmarks than what existed publicly, because high-quality document datasets are scarce. Reducto invested in rigorous data pipelines, expert input, and diversified sampling to iterate quickly and evaluate against alternatives.
- •Lack of high-quality public document datasets slowed the field
- •Built an internal “data engine” and extensive pipelines
- •Sourced diverse document data to test long-tail performance
- •Fresh perspective (vision-first) avoided legacy heuristic approaches
- 12:45 – 13:42
Beyond PDFs: expanding to spreadsheets, images, and slides without forking pipelines
Reducto broadened format support as customers repeatedly asked for more endpoints to avoid maintaining parallel ingestion systems. They balanced expansion with maintaining a high accuracy bar, aiming to be a single ingestion layer across document modalities.
- •Now supports spreadsheets, images, documents, and slides
- •Customer pain: having to fork ingestion pipelines per format
- •Expansion driven by repeated requests (e.g., spreadsheet endpoint)
- •Accuracy bar remained a constraint while broadening scope
- 13:42 – 14:31
Why Reducto becomes core AI infrastructure: moving faster with new models and better reasoning
They position Reducto as foundational infrastructure for AI apps and agents, especially when teams previously built ingestion in-house to hit quality targets. Offloading ingestion helps teams adopt new models faster and spend effort on reasoning and product features instead.
- •Core value: eliminate ingestion as the blocker to quality
- •Helps teams keep up with new models and evolving best practices
- •Lets builders focus on post-processing, reasoning, and product UX
- •Applies to startups through large enterprises
- 14:31 – 15:34
Hiring for craftsmanship: detail-oriented engineers who obsess over data quality
Following their Series A announcement, they describe hiring across engineering roles, including ML and product engineering. They emphasize scrappiness and deep attention to detail—highlighting that data quality directly determines customer results.
- •Hiring across engineering roles (ML, product, general engineering)
- •Prefers scrappy builders with startup/founder-like mindset
- •Quality comes from meticulous inspection and iteration
- •Principle: output quality is constrained by input data quality
