---
title: "Priyansh Singhal (@priyanshsinghal) | Hashnode"
description: "Full-stack & AI engineer, writing about RAG, LLM agents, and GenAI. Read the latest articles by Priyansh Singhal on Hashnode."
url: "https://hashnode.com/@priyanshsinghal"
---

@priyanshsinghal

# Priyansh Singhal

@priyanshsinghal · New Delhi, India · Joined September 2026

Full-stack & AI engineer, writing about RAG, LLM agents, and GenAI.

### [I Showed My Coding Agent Its Own Bug. It Copied the Bug Twice.](https://priyanshsinghal.hashnode.dev/i-showed-my-coding-agent-its-own-bug-it-copied-the-bug-twice?utm_source=hashnode&utm_medium=feed)

A blind "try again" fixed the bug on the first extra attempt. Telling the model exactly what it got wrong did not help at all. My local model wrote a next-permutation function that failed 3 of 5 test

9 min read · [priyanshsinghal.hashnode.dev](https://priyanshsinghal.hashnode.dev/i-showed-my-coding-agent-its-own-bug-it-copied-the-bug-twice?utm_source=hashnode&utm_medium=feed)

[1d ago](https://hashnode.com/posts/i-showed-my-coding-agent-its-own-bug-it-copied-the-bug-twice/6afcfa3ce9f8e28b98eba7a3)

### [My LLM Judge Caught 1 of 41 Wrong Answers. One Line Fixed the Rest.](https://priyanshsinghal.hashnode.dev/my-llm-judge-caught-1-of-41-wrong-answers-one-line-fixed-the-rest?utm_source=hashnode&utm_medium=feed)

Asking the judge to solve the problem itself before grading turned a rubber stamp into a real check.

9 min read · [priyanshsinghal.hashnode.dev](https://priyanshsinghal.hashnode.dev/my-llm-judge-caught-1-of-41-wrong-answers-one-line-fixed-the-rest?utm_source=hashnode&utm_medium=feed)

[Nov 12](https://hashnode.com/posts/my-llm-judge-caught-1-of-41-wrong-answers-one-line-fixed-the-rest/6af54e187062cb28d849e4c7)

### [My Agent Had the Right Tool in a 50-Tool Menu. It Used the Wrong One.](https://priyanshsinghal.hashnode.dev/my-agent-had-the-right-tool-in-a-50-tool-menu-it-used-the-wrong-one?utm_source=hashnode&utm_medium=feed)

Raw accuracy fell 25 points at 50 tools, but a chance-corrected read tells a very different story.

9 min read · [priyanshsinghal.hashnode.dev](https://priyanshsinghal.hashnode.dev/my-agent-had-the-right-tool-in-a-50-tool-menu-it-used-the-wrong-one?utm_source=hashnode&utm_medium=feed)

[3](https://hashnode.com/posts/my-agent-had-the-right-tool-in-a-50-tool-menu-it-used-the-wrong-one/6aebe566e425433ca7cea95c) [Nov 5](https://hashnode.com/posts/my-agent-had-the-right-tool-in-a-50-tool-menu-it-used-the-wrong-one/6aebe566e425433ca7cea95c)

### [I Reproduced OpenAI's Agent Safety-Alert Evasion on a Free Model](https://priyanshsinghal.hashnode.dev/i-reproduced-openai-s-agent-safety-alert-evasion-on-a-free-model?utm_source=hashnode&utm_medium=feed)

10 of 25 runs disabled their own alert to follow a conflicting instruction, and none lied about it

9 min read · [priyanshsinghal.hashnode.dev](https://priyanshsinghal.hashnode.dev/i-reproduced-openai-s-agent-safety-alert-evasion-on-a-free-model?utm_source=hashnode&utm_medium=feed)

[Oct 28](https://hashnode.com/posts/i-reproduced-openai-s-agent-safety-alert-evasion-on-a-free-model/6ae1ed92646191b43a282b12)

### [My LLM Router Spent 8x More and Matched the Cheap Model Exactly](https://priyanshsinghal.hashnode.dev/my-llm-router-spent-8x-more-and-matched-the-cheap-model-exactly?utm_source=hashnode&utm_medium=feed)

Two opposite effects from one confidence prompt canceled out, and only checking the math caught it.

11 min read · [priyanshsinghal.hashnode.dev](https://priyanshsinghal.hashnode.dev/my-llm-router-spent-8x-more-and-matched-the-cheap-model-exactly?utm_source=hashnode&utm_medium=feed)

[Oct 21](https://hashnode.com/posts/my-llm-router-spent-8x-more-and-matched-the-cheap-model-exactly/6ad8b2943242cf8246919e35)
