Ravikumar M. N.
Menu
Résumé
Selected work

Case study

Agentic document intelligence, from PDFs to structured data and answers

Docling parsing and LLM extraction turn PDFs, forms, and reports into JSON, a knowledge graph, and grounded QA, at under 3s per page.

Role
AI / ML Engineer
Context
Zyoin · Acranton · projects
Core stack
Docling, LLMs, Gemini API
ingestion per page
<3s
grounded QA response
1.5s
healthcare form throughput
+30%

Problem

Most enterprise knowledge sits in documents nobody can query: PDFs, scanned forms, and long reports. The goal was to make that content usable in two ways at once, as structured data that downstream systems can consume and as answers people can get by asking questions.

Constraints

  • Messy inputs. Layouts, tables, forms, and handwriting vary widely between documents.
  • Throughput. Collections are enterprise-scale, so ingestion has to stay fast per page.
  • Grounded answers. Answers and summaries must come from the documents themselves.
  • Healthcare data. Some of it was medical. Extraction had to be fast enough for clinicians to use.

Approach

Parse structure before reasoning

Each document goes through Docling first, which recovers layout, tables, and text. The LLM then works on structured content instead of raw PDF bytes. That is cheaper and more reliable than asking the model to also work out the page layout.

Agent-driven extraction and classification

An extraction agent classifies each document and extracts fields into structured JSON and a knowledge graph. The same content is chunked and embedded into a vector index for semantic search.

Conversational QA and summarization

A RAG pipeline over the vector index answers questions and summarizes across ingested documents, with 1.5s grounded QA latency on enterprise-scale collections.

Agentic document intelligenceIngestion: documents (PDFs, forms, reports) are parsed by Docling, then an LLM extraction agent classifies and extracts them into structured JSON, a knowledge graph, and a vector index, in under 3 seconds per page. Query: a QA and summary agent retrieves context from the index and returns a grounded answer in about 1.5 seconds.INGESTION · <3s PER PAGEQUERY · 1.5s GROUNDED QAtop-k contextDocumentsPDFs · forms · reportsDocling parserlayout · tables · textLLM extraction agentclassify · extractStructured JSONKnowledge graphVector indexchunks · embeddingsUser questionQA & summary agentretrieve · groundGrounded answer1.5s end-to-end
Fig.The ingestion and query planes. All components are from the résumé; the edges between them are simplified.

The same pattern in healthcare

  • Form extraction (Acranton). A Gemini API–based extractor converts unstructured healthcare documents into structured data in under 3 seconds, raising throughput 30%.
  • Handwritten prescriptions. Gemini Flash converts handwritten prescriptions into JSON within 5 seconds, making the data accessible to clinicians.

Results

document ingestion per page
<3s
grounded QA response latency
1.5s
healthcare form → structured data; throughput gain
<3s · +30%
handwritten prescription → JSON
<5s

Stack

Parsing
  • Docling
Models
  • LLMs
  • Gemini API
  • Gemini Flash
Serving
  • FastAPI
  • Vector DB
  • RAG