AI Researcher · Engineer
Hey there, I'm
I research and build retrieval-augmented generation that helps language models read, retrieve and reason over real-world documents.
UA
Years in AI & ML
Publications
Citations
First-author papers
01 — A little about me
I'm an AI researcher and engineer working on retrieval-augmented generation and document understanding for large language models.
For more than eight years I've worked on applied AI. I started as an AI scientist at Course5i, then moved into machine learning engineering at yellow.ai, where I now lead the engineering of the Document Cognition product.
My research looks at how documents should be chunked, ingested and retrieved so language models give grounded, accurate answers. I'm based in the United Arab Emirates and always happy to talk about RAG, document AI and LLMs.
Connect on LinkedInHow documents should be chunked, ingested and retrieved so LLMs give grounded, accurate answers.
Multimodal ingestion pipelines that turn PDFs, web pages and tables into retrieval-ready knowledge.
Senior AI/ML Engineer at yellow.ai, leading the Document Cognition product.
02 — Publications
A GraphRAG system where deleting a document is an exact database operation with no model calls. Across MuSiQue, 2WikiMultiHopQA and HotpotQA it builds the cheapest index of the systems compared: 26–31% cheaper than HippoRAG 2 and 3–6× cheaper than LightRAG and fast-graphrag.
Uses multimodal document understanding to decide how documents are split into chunks for retrieval-augmented generation, so structure like tables and multi-page sections isn't broken apart.
Retrieval-aware ingestion for enterprise documents, normalising PDFs and converting them to multimodal Markdown before chunking.
A retrieval-aware chunking approach for web content, aimed at making RAG systems more efficient and cheaper to run.
Examines where large language models fall short when following instructions.
Adds context to tables before passing them to language models, so the models can summarise tabular data more efficiently.
Earlier work
International Conference on Man and Machine Interfacing (MAMI), 2015 · Cited by 2
Don't just take
What researchers and practitioners have said about the work on X, Hugging Face, Medium and YouTube.
“Vision‑guided chunking leverages LMMs to chunk PDFs by preserving visual & cross‑page context.”
“Built on the W-RAC framework, I observed a reduction in average token usage from ~2500 to ~345 tokens.”
Featured as #437 in the Papers Explained series: a 10-minute deep-dive into the method and results.
“Glad to see more papers focusing on multimodal RAG.”
An independent video walkthrough of the paper and its approach to chunking complex PDFs.
Another independent video breakdown of the paper.
03 — Where I've been
United Arab Emirates · Full-time
Bengaluru, India · Full-time
Bengaluru, India · Full-time
Education
Manipal Academy of Higher Education
Miracle Educational Society Group of Institutions
04 — Beyond the desk
Away from the keyboard, I'm usually somewhere new. 7 countries so far, and counting.
Photo gallery coming soon. Until then, the trips live on Instagram.
Follow @uday_allu05 — Say hello
Whether it's a research collaboration, a role, or just a question about my work, my inbox is open.