Adit Abraham
Related people
Episodes
Notes
a16z · The Missing Infrastructure for AI Agents: Documents | Reducto’s…
-
Can Agents Do Healthcare Work Without Reliable Document Parsing?
Jennifer Li asks how accuracy changes whether customers will trust an agent with their documents. Adit Abraham says a document error in healthcare is not a small miss. One large customer gives doctors a search tool over physical packets that are inches thick. A misread can make a history look absent, and an overworked doctor who trusts the tool may then treat the patient as someone who maybe does not need the drug. If an agent made that decision, everything after it would depend on inputs you can trust.
On those harder packets he still sees cases below 50 percent accuracy, where the best the tool can do is approximate where information sits and a person has to look. One patient's packet can
-
Do Frontier Models Silently Drop Facts From Long Documents?
Adit Abraham says frontier models were more precise than many specialized document parsers on a long-extraction benchmark: when they pulled a fact, it was more likely to be correct. The failure was recall. They silently dropped information, and that miss is hard to catch once the pipeline is in production.
A single model call is the wrong shape for a long document. The file has to be broken into parts. The Deep Extract harness is not tied to one model. It gives the model tools to review its own output and to apply checks, for example that the lines on an invoice add up to the total, so it knows when to look again. He says that review loop is what got the extraction to near-perfect accuracy
-
Why Won't the Big Model Labs Solve Document Parsing?
Jennifer Li asks why the big model labs, already focused on knowledge work and on vision, will not simply solve document parsing. Adit Abraham says this work is tangential to intelligence. Reducto is not competing with the labs on reasoning quality, and the range of problems runs deeper than one model. A PDF can contain a folder of other PDFs. Rendering a Microsoft Word file faithfully still lacks good open-source tools, which is why LibreOffice often falls short.
He believes the labs are in an arms race to create superintelligence, and that document infrastructure is what makes that useful. Focus is the other advantage. A vision encoder for ChatGPT has to take a selfie and turn it into a S
-
Why Do Document Mistakes Compound Across a Dozen Files?
Adit Abraham says a loan packet is more than reading the page. The job splits the files, extracts individual fields, and may edit the application at the end, and each of those steps has to be accurate. Across a dozen files, 90 percent on the individual pages is not acceptable, because those mistakes compound.
Jennifer Li says a lot of traditional OCR still lands around 70 to 80 percent, short of the high 90s. For parsing itself, Adit says the vast majority of use cases can already reach the high 90s, aside from corner cases such as handwriting he cannot read. The newer work is a harness around the model. Deep Extract is where he finds an agent can beat humans at structured extraction on man
-
What Should You Check First in a Document Benchmark?
Adit Abraham says the first thing to check is whether the distribution of data matches the work you want done. Simple documents often need only a simple pipeline, and a generic benchmark is the wrong yardstick. Strict scoring rules are a mistake, because the same information can be written faithfully two different ways and should not be penalized.
Benchmarks in the AI industry have become marketing, so everything looks state-of-the-art and it is hard to know what is good in your own context. A public benchmark they tested shows an egregious version. Take the same documents, tilt each page one degree, and add a little Gaussian blur. The provider that published the benchmark gets way worse, w
-
Why Did Coding Tools Reach a Billion in Revenue First?
Adit Abraham says coding tools reached a billion dollars in annual revenue while professional-services AI has not, even though healthcare, insurance, legal, and finance are full of valuable paper-heavy work. The gap is second-order problems, not a lack of work worth doing.
In 2023, code generation tools were mostly code completion and only a fraction of that market. A friend came to dinner with a shaved head after the Cursor team shaved their heads to celebrate a million in annual revenue, and the milestone felt big. Many companies in the code generation market are now past a billion in annual revenue. Professional services markets have not shown the same jump. He wants to help customers so
-
How Does Reducto Read a Messy Document?
Adit Abraham breaks document reading into a few epochs, ending with Reducto R1, one unified model that reads a page from end to end. Bounding boxes stay in place, the result stays consistent, and the model captures everything on the page without the accuracy and latency trade-offs older setups had to accept.
PDF software started as a way to print. People in the 1990s wrote drivers so every character would appear on the page. OCR companies later templated a steady form such as a W-2, pinning the pixels where a name sits and cleaning the text afterward. Templates were enough for a lot of consistent pipelines. In 2023, nobody thought about complexity before uploading a document to ChatGPT. The