By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
At a glance
WHAT IT’S REALLY ABOUT
Reducto turns messy enterprise documents into accurate LLM-ready structured data
- Reducto converts complex, messy enterprise documents (claims, health records, financial statements) into clean structured data optimized for LLM workflows like RAG and summarization.
- The company emerged from an initial long-term LLM memory project after customers repeatedly asked for better handling of uploaded files, revealing PDF ingestion as a major bottleneck.
- Their key technical shift was treating document parsing as a computer-vision problem—“reading like a human”—rather than relying on brittle heuristics and format-specific rules.
- Customers report major downstream gains from simply swapping ingestion providers, including large LLM accuracy jumps and formerly impossible features such as reliable processing of scanned and handwritten documents.
- Reducto rapidly won large enterprise deals (including a Fortune 25) by outperforming internal document-processing teams and investing heavily in high-quality evaluation data, benchmarks, and iteration pipelines.
IDEAS WORTH REMEMBERING
5 ideasDocument ingestion quality is a primary limiter of LLM product performance.
Customers see significant end-task accuracy improvements (reported up to ~30%+) by improving extraction/layout/chunking before the model ever reasons, because the LLM can only work with what it’s given.
PDF processing is “schlep work” that many teams underestimate until it blocks shipping.
Even strong AI teams end up spending disproportionate time on parsing and post-processing because off-the-shelf tools often fail on real-world variability, yet few companies want to specialize in it.
A vision-first approach can generalize better across the long tail of document formats.
Reducto reframes parsing as understanding visual structure (spacing, hierarchy, layout) like a human reader, avoiding brittle rule sets tied to specific templates or PDF standards.
Modern VLMs unlock extraction cases that traditional OCR struggles with.
Handwriting, scanned documents with missing metadata, checkboxes, and unusual markings (e.g., highlighted/circled table cells) become tractable when combined with CV layout models and targeted model orchestration.
“Be the ingestion team” is a compelling wedge into the AI application stack.
By abstracting ingestion, Reducto lets application teams focus on product logic and reasoning layers while still reaching the quality bar needed for enterprise deployments.
WORDS WORTH SAVING
5 quotesWe help people take their really, really complicated documents… and turn that into clean structured data… primarily LLM-based use cases like RAG and summarization.
— Reducto founder
None of these AI application layer companies want to be PDF processors… we try to be the ingestion team for the companies that we work with.
— Reducto founder
We turned PDF processing… into a computer vision problem… parse and understand these documents the way humans do.
— Reducto founder
People will often see end LLM accuracy improvements just from swapping the ingestion provider.
— Reducto founder
The quality of your data is the quality of your end outputs and results.
— Reducto founder
SPEAKERS
Diana Hu
hostHost/interviewer on YC Root Access.
High quality AI-generated summary created from speaker-labeled transcript.
