Adit Abraham

Adit Abraham is the co-founder and CEO of Reducto, which builds document-processing infrastructure for AI applications. His work addresses a consequential weakness in otherwise capable systems: an agent can reason convincingly from a financial table it has read incorrectly. At Reducto, Adit combines product development and direct customer work with a technical position that shapes the platform: document processing must preserve what the source says, including its mistakes, while making its contents usable for retrieval and reasoning.

From a memory product to document parsing

Adit studied computer science at MIT and conducted machine-learning research at the MIT Media Lab. His early entrepreneurial projects included Sidewalk, a Shopify plugin that recommended products from other stores during checkout. He met his future co-founder, Raunak Chowdhuri, in a graduate machine-learning course at MIT. Their partnership paired Adit’s product and customer-development work with Raunak’s technical focus, as described in Reducto’s path to product-market fit.

Before Reducto, Adit worked as a product manager at Google on YouTube advertising and Search. He valued advertising’s clear commercial objectives and measurable results; his Search work involved more coordination and a less immediate connection between his work and its impact.

Adit and Raunak founded Reducto in 2023 and joined Y Combinator’s Winter 2024 batch. Their original product, Remembrall, gave language models long-term memory. Conversations with prospective customers kept returning to an earlier obstacle in the workflow: getting usable information out of uploaded files. The founders had encountered the same difficulty in their own application. A document-processing experiment attracted stronger demand than the memory product, prompting their pivot to parsing.

An early playground invited prospective customers to upload documents that had defeated other tools. This let the founders demonstrate an improvement on the customer’s own difficult table or scan. Adit initially concentrated on startups in finance, insurance, and legal work, where accuracy mattered and teams could integrate quickly. Manual onboarding exposed failures and requirements that a polished demonstration could miss. In his account of founder-led customer development, Adit explains why he retained direct involvement in sales and customer success as sources of product knowledge.

Enterprise customers brought additional demands. Reducto’s first Fortune 10 sales process lasted 154 days and included technical evaluations, acquisition discussions, procurement negotiations, and an air-gapped deployment. The founders’ account of the deal emphasizes the role of an internal customer champion and the delivery work that continued after signing. Showing that Reducto could read the documents was one requirement; making it work within the customer’s deployment and procurement constraints was another.

Preserve the source, then make it usable

Adit draws a precise distinction between correcting a reading error and changing a document. A parser should repair a misread zero or punctuation mark. If a financial statement contains an incorrect total, however, it should preserve the printed number. A model that recalculates the total has altered the evidence on which later decisions depend.

In his discussion of document intelligence, Adit explains how that distinction informs Reducto’s agentic OCR: an initial parse is followed by verification and targeted correction. Regenerating the entire text can introduce new errors when a model substitutes something more plausible for what is actually on the page. As agents make successive decisions, those substitutions can propagate beyond the original document.

Several other choices define his approach:

• Match the representation to its consumer. Markdown can represent simple tables efficiently, while HTML can preserve relationships encoded by merged cells. But an embedding model may struggle to connect a question about revenue over time with a block of tags and numbers. Adit advocates keeping the structured table for reasoning and creating a separate natural-language representation for retrieval. The same information serves two different tasks without forcing one format to do both poorly.

• Use vision models selectively. Adit favors conventional computer vision for tasks such as layout detection and vision-language models where semantic interpretation matters, including handwriting. Small layout models can run efficiently at scale; larger models can concentrate on difficult regions. For line charts, Reducto’s team has used agents with code interpreters and plotting tools to reconstruct data, visualize the result, and check it repeatedly. The team’s chart-extraction research also examines the limits of visual verification and the use of specialized models for axes, legends, and data series.

• Evaluate the steps as well as the final answer. Parsing, retrieval, formatting, and agent performance can fail independently. A faithful transcription is insufficient if retrieval supplies the wrong passage. Adit advocates evaluating each stage and monitoring production documents, which can differ from benchmark examples. Reducto’s RD-TableBench, with 1,000 manually annotated examples of difficult tables, is one of the team’s contributions to evaluating document understanding.

Adit also treats classification and splitting as substantive parts of the workflow. A packet containing hundreds of pages may combine several document types or unrelated sections. Sorting and routing that material lets a model concentrate on the relevant task rather than spend its effort disentangling the packet. He describes file-system and CLI interfaces through which agents choose document tools as needed, with readable content kept distinct from metadata such as citation bounding boxes. This gives agents alternatives to sending every file through the same predetermined sequence.

From parsing accuracy to customer workflows

Reducto’s collaboration with Blueflame AI and Datasite shows how these choices affect an application team. The relationship began while Blueflame was independent and continued after Datasite acquired it in July 2025. Bringing parsing into Datasite’s virtual data rooms meant processing confidential transaction records within demanding security and deployment requirements. Adit’s team worked with Datasite to establish how the content would remain protected during processing.

In their conversation about the integration, Adit asks Datasite’s global CTO and Blueflame co-founder, Raj Bakhru, about the decision to buy parsing infrastructure. Raj describes a private equity customer reporting a wrong answer caused by a vision model matching a pie slice to the wrong shade of blue in its legend. Investigating such extraction failures consumed engineering time that his team wanted to spend on search, GraphRAG, interfaces, and agents.

Raj reported that Reducto parsing improved recall by 22% on Datasite’s internal VDR Bench evaluation. The benchmark used questions with known answers against a virtual data room containing hundreds of documents. That customer-reported result concerns a particular workflow; it illustrates how improving ingestion can change retrieval performance and where an application team spends its effort.

Under Adit and Raunak, Reducto expanded from parsing into structured extraction, splitting, classification, and document editing. The founders announced a $75 million Series B in October 2025, bringing total funding to $108 million. In September 2026, the company introduced r-1 in preview, a parsing model designed to handle layout, tables, formatting, and citations in a single pass. That development adds a unified-model approach to the multi-stage methods Adit has described, while keeping the practical requirement intact: downstream systems need both the document’s contents and the structure that gives those contents meaning.