Who is Munna Suprathik?

Munna Suprathik is a Generative AI Engineer based in Bengaluru, India. He builds production RAG systems, multi-agent pipelines, and LLM-powered apps, and works across Claude, GPT, and Gemini.

What does Munna Suprathik do?

He builds RAG systems, AI agents and chatbots, LLM-powered web apps, AI integrations into existing products, and Shopify stores. His deeper focus is agentic systems and automation with LangGraph, MCP, and A2A.

Is Munna Suprathik available for hire?

Yes. He takes on freelance and contract work. The fastest way to reach him is email at hello@munnasuprathik.com or LinkedIn.

What has Munna Suprathik built?

Live products include Tend, an agentic task app; Margam AI, a RAG study-abroad platform; Dissect, a viral-content analysis SaaS; and Quiccle, an AI discovery hub.

What is Munna Suprathik's tech stack?

Python, LangChain, LlamaIndex, LangGraph, vector search, Pinecone, Cohere Rerank, HyDE, parent-child chunking, LLM-as-judge evaluation, function calling, MCP, A2A, and LoRA and PEFT fine-tuning. He builds across Claude, GPT, and Gemini.

Notes · Munna Suprathik

When I Trusted the LLM to Grade Itself

By Munna Suprathik, Generative AI Engineer.

I wanted to test hundreds of outputs without reading each one, so I did the obvious thing and had an LLM grade another LLM. For about a day it felt like I had cheated the universe.

Then I spot-checked. The judge was handing 9 out of 10 to answers that were flatly wrong. It had opinions, it had confidence, and it was hallucinating exactly like the model it was supposed to police.

A grader with no ground truth is just a second guesser in a lab coat.

Scoreboard of numbers — Confident scores, nothing underneath them.

What actually made it trustworthy was treating the judge like code, not an oracle. I gave it a rubric with concrete pass/fail criteria instead of asking "is this good." I forced it to quote the exact span it was judging so it could not vibe its way to a score. And I hand-labeled maybe forty examples as a fixed answer key, then measured the judge against me. If it drifted from the human labels, the judge got fixed, not the model.

LLM-as-judge is genuinely useful for catching regressions at scale. It just moves the hallucination up one floor if you forget to audit it, and I forgot.