Hello guys :-

Here's my portfolio

portfolio.nirvan.dev

I want genuine suggestions about my portfolio

It will help me a lot

Hey everyone,

I’ve been using AI assistants (Claude/Cursor) for a while, but I kept running into two annoying issues:

1. Package hallucinations: The AI suggests an import for a package I don’t even have in my package.json.
2. Version/Method hallucinations: It suggests a method like prisma.user.findFirstOrThrow() when I’m on an older version of Prisma where that doesn't exist yet.

To fix this, I built ctxai—a Model Context Protocol (MCP) server that makes the LLM "version-aware."

How it works:

It injects a "version fingerprint" of your actual environment into the AI’s context. Before the AI gives you a code snippet, it runs a validation layer to check if the code actually works with your installed versions.

The tools it adds to your AI:

• get_project_context: Scans your root and maps every installed package/version (Node & Python).
• validate_suggestion: A 3-layer validator that catches missing packages and hallucinated methods.
• get_package_docs: Fetches live metadata from npm/PyPI to help the AI self-correct.

Current Tech Stack Support:

• Node.js: Reads package.json and uses TypeScript definitions for method checks.
• Python: Reads requirements.txt/pyproject.toml and handles tricky import-to-pip name mappings (like PIL → Pillow).

I also included a benchmark suite of 28 test cases to ensure it’s actually catching errors without too many false positives.

I’m a student and still learning the ropes of MCP, so I’d love some feedback on the architecture or the validation logic!

GitHub: https://github.com/Nirvanjha2004/ctxAI-MCP-tool-to-reduce-hallucinations/tree/main

Would love to know: Are there any specific libraries where you find AI hallucinations to be the most frequent? I want to add more "mock API surfaces" to the benchmark.

I am from mechanical from tier 3 cllg, i gave amazon OA on 4th for sde intern position , but even after passing all the test cases and doing the behavioral MCQs correctly , I was not shortlisted for interview and some guys who solved 1 and half question coding got selected ( they from cse and ece ) , why is it happening ?? And is there any chance I can get interview link ??