VP Vinod Pungle

Published Work

Writing and open-source work, grounded in shipped systems.

I write about the patterns I've actually built and delivered — reliability in RAG systems, knowledge graphs, and what it takes to move GenAI from demo to production.

Featured article

From the blog

Stop Hallucinations in RAG: Why Knowledge Graphs Change Everything

vinood.hashnode.dev

GraphRAG · LLM reliability · Enterprise AI

Pure vector retrieval breaks down in high-stakes domains: semantically similar isn't the same as factually connected, and that gap is where hallucinations live. This piece makes the case for knowledge-graph-grounded retrieval — multi-hop relationships, explicit dependencies, verifiable answers — drawn from building exactly this architecture for enterprise audit content and in my open-source hybrid RAG system.

GraphRAGHallucination mitigationNeo4jRAG architecture
Read on Hashnode

Blog

More writing

New articles on agentic AI, retrieval architecture, and AI delivery are published on my Hashnode blog.

Visit the blog

Open source

Code as publication

My repositories are the working companions to the writing — most notably the hybrid RAG capstone that implements the knowledge-graph retrieval patterns discussed in the featured article.

Browse featured repositories

Coming soon

The writing pipeline

Articles in progress — each grounded in systems I've shipped or programmes I've run. Follow the blog or LinkedIn to catch them as they land.

Delivering AI programmes · AI-enabled delivery

The delivery leader's lens

  • Delivering Programmes with AI — GenAI and agents in estimation, reporting, risk, and backlog intelligence
  • Estimating the Un-estimatable — backlogs and budgets for GenAI discovery phases
  • What SAFe PI Planning Looks Like When the Deliverable Is an AI Agent
  • The AI Project Is Not a Software Project — 12 years of Agile meets ML delivery
  • The Retraining Budget — model maintenance as a first-class programme line item

RAG reliability & knowledge graphs

Trustworthy retrieval

  • Explainability When the Regulator Asks — grounding, citations, and knowledge graphs in audit-grade AI
  • From Vector Search to Knowledge Graphs — architecting hybrid RAG in production
  • How We Measured RAG Quality Before the Client Did — evaluation in practice
  • Multi-hop Reasoning in Audit-grade Systems — where pure RAG fails
  • Your RAG System Rots Quietly — detecting content drift and retrieval decay

LLMOps · observability · lifecycle

Running AI in production

  • LLMOps Is Not MLOps — what changes when the model isn't yours
  • Observability for LLM Systems — what to log when the output is a paragraph
  • Traceability End to End — from an answer back to document, chunk, embedding version, and prompt
  • When to Retrain, When to Re-prompt, When to Re-index — a decision framework
  • Blue-Green for Embeddings — re-indexing a production vector store without downtime
  • Deploying RAG Behind API Management — rate limits, fallbacks, and cost guardrails on Azure

Agentic engineering

Agents that ship

  • Orchestrating Specialised Agents with Azure Durable Functions — an enterprise pattern
  • Guardrails and Evals — lessons from Google's AI Agents Intensive
  • Airflow for RAG Pipelines — orchestrating ingestion, embedding, and index refresh
  • DVC + DagsHub in Anger — data versioning that survives a real team
  • From docker compose up to Terraform — the promotion path of an ML prototype

Live applications

Deployed demos

Hosted demos of selected projects are coming soon. Meanwhile, the hybrid RAG system runs locally with one command — full setup instructions in the repo.

Run it yourself

Hybrid RAG — Streamlit demo

Clone the repository, bring the stack up with Docker Compose, and launch the Streamlit UI to query documents through hybrid vector + graph retrieval.

docker compose upstreamlit run ui/app.py
Get the code

Watch this space

Hosted demos — coming soon

Live, hosted versions of the agentic and RAG demos are planned. Follow the blog or GitHub for release announcements.

Follow on GitHub

Talking through a retrieval or agent architecture problem?

I'm always glad to compare notes on RAG reliability, knowledge graphs, and agentic patterns.

Get in touch