Sean Sodha

Sean Sodha is a deep learning senior product manager at NVIDIA, responsible for multimodal retrieval and document ingestion for NeMo Retriever. His work addresses a practical problem in enterprise AI: making tables, charts, images, and text searchable while preserving the relationships that give their contents meaning.

Before joining NVIDIA, Sodha ran his own AI venture and worked at IBM Watson. His education combines engineering and business: he holds a bachelor’s degree in electrical engineering from Purdue University, a master’s degree in engineering from Cornell University, and an MBA from the Wharton School. His public author biography documents this background and his NVIDIA product responsibilities.

From PDF extraction to multimodal retrieval

Sodha co-authored NVIDIA’s August 2024 multimodal PDF extraction blueprint with Tanay Varshney, Annie Surla, and Nicola Sessions. The guide explains how specialized extraction models, embeddings, and reranking fit together in a retrieval-augmented generation pipeline: a system that finds relevant material in a document collection and supplies it to a language model answering a question.

The workflow first separates a PDF’s content into different modalities. Object detection identifies charts and tables; chart transcription and optical character recognition extract their contents and metadata. Table extraction preserves reading order. The resulting material is filtered, divided into chunks, embedded, and stored for search. When a question arrives, vector search retrieves candidate passages, a reranker refines the results, and a language model uses the selected context to generate an answer. This division of work lets each stage handle a specific problem rather than treating the entire PDF as undifferentiated text.

In February 2026, Sodha co-authored a Nemotron document-processing tutorial with Chia-Chih Chen, Moon Chung, and Nave Algarici. It extends the approach to embeddings and reranking that can process images alongside text. The tutorial uses the World Bank’s Peru 2017 Country Profile, whose narrative, charts, and dense appendix tables exercise different extraction capabilities. Tables remain in Markdown, charts are cropped as images, and extracted items retain page metadata. The pipeline can therefore retrieve visual evidence directly and carry its source location into answer generation.

Practical choices in document AI

These co-authored guides explain several decisions that shape retrieval quality and deployment:

• Preserve relationships inside tables. Flattening a table into text can separate a value from the row or column that explains it. The 2026 tutorial uses Markdown to retain those relationships, giving downstream models a clearer representation of what each number means.

• Use search to narrow the field, then judge relevance. Embeddings make similarity search efficient across a large collection. A cross-encoder reranker then evaluates the question together with each shortlisted item and, where available, its image. Applying this more expensive step to a shortlist allows closer relevance assessment without scoring every document against every question.

• Balance precise citations with broader context. Page-level extraction makes it easier to identify where an answer came from; larger document units preserve more narrative context. Chunk size creates a related tradeoff: smaller chunks can support precise retrieval while omitting surrounding information needed to interpret a passage. Retaining source metadata helps users check the evidence selected for an answer.

• Match deployment to the workload. The 2026 tutorial distinguishes lightweight library execution for development and small collections from container deployments using Redis or Kafka to distribute larger workloads. The earlier blueprint explains how containerized NVIDIA NIM microservices, standard APIs, and Helm charts support deployment and scaling. Together, the guides connect extraction and retrieval mechanisms to the infrastructure needed to process enterprise document collections.