Writing · Ardya Dipta Nandaviri

Featured deep dive

Architecture first, vendor choices second.

Production RAG · Long-form guide

From 100,000 documents to grounded answers

A practical design for ingesting mixed PDFs, scans, images, tables, and diagrams—then retrieving them with metadata-aware BM25 and semantic search, reciprocal rank fusion, reranking, and selective GraphRAG.

RAGSystem designHybrid retrievalGraphRAG

Read the article Play the animation

100K

mixed documents enter the pipeline

1.5Mpages

2Mchunks

3retrievers

All writing

Search by title, summary, or topic.

4 posts Static HTML · no tracking

ArticleLong-form guide

From 100,000 Documents to Grounded Answers

A production-oriented RAG architecture: tiered parsing, canonical document IR, metadata and ACL filtering, BM25 plus HNSW, RRF, reranking, evaluation, and selective GraphRAG.

RAGHybrid retrievalGraphRAG

Interactive12-step animation

Inside a 100,000-document RAG knowledge base

Play, pause, and scrub through the entire production RAG pipeline—from raw mixed documents to a streamed, cited answer.

InteractiveRAGSystem design

Medium articleGenAI

How Google Gemini can help create a “Mike Ross” chatbot agent

Building a document-querying assistant with Google Gemini, FAISS, and LangChain for fast, grounded access to organizational knowledge.

GeminiFAISSLangChain

Medium articleReinforcement learning

Teaching AI to understand us with RLHF

An introduction to fine-tuning language models with reinforcement learning from human feedback, from preference data to improved model behavior.

RLHFLLMsFine-tuning

About the author

Ardya Dipta Nandaviri

Senior AI/ML Consultant at AWS Professional Services with 10+ years of experience building production AI/ML systems. Previously Head of Data Science at Kalbe Farma and Lead Data Scientist at Gojek; Carnegie Mellon Robotics alum based in Singapore.

View portfolio Follow on LinkedIn Medium

Ideas are useful when they survive production.