Pulkit Jain

Amplifying AI - Builder, Hustler, Thought Leader | AI Immune Systems

[Pulkit Jain @pulkitskingdom]

Repo - https://t.co/yNAoO1meMX
Python Package - https://t.co/U4risyjW7z

[Pulkit Jain @pulkitskingdom]

1/6 Two researchers proved this year that AI hallucination is mathematically impossible to fully eliminate.

So I stopped trying to eliminate it.

I built a way to detect it instead.

Groundguard — open source, available now. 🧵

[Pulkit Jain @pulkitskingdom]

6/6 Built over several weeks.

Tested against real documents across multiple providers.

Every design decision reviewed adversarially before a line of code was written.

MIT licensed.

GitHub + PyPI 👇

[Pulkit Jain @pulkitskingdom]

8/8 Three pillars. None sufficient alone.

The generator misses things. The critic catches different things. Your process discipline catches what neither sees.

Full breakdown → https://t.co/RZws2tGAVQ

[Pulkit Jain @pulkitskingdom]

1/8 I had no engineering team. What I had instead: a method. Three pillars that shipped a production-grade open source library. 🧵

[Pulkit Jain @pulkitskingdom]

7/8 During Gemini testing: 25–28 API calls per test.

Each test should make exactly one.

A test making 25 calls when it should make 1 is telling you something about the architecture — not just the test.

[Pulkit Jain @pulkitskingdom]

6/6 In production, most verifications never touch an LLM.

That's not a feature.

That's the only way this architecture works at enterprise scale.

Full writeup on LinkedIn → https://t.co/RFOxN5BoUl

[Pulkit Jain @pulkitskingdom]

1/6 I started with one function.

verify(claim, sources) → VERIFIED / CONTRADICTED / UNCERTAIN

Simple. Useful. Immediately incomplete.

[Pulkit Jain @pulkitskingdom]

5/6 Still unsolved: cost.

$0.015 per LLM call × 10,000 daily verifications = $150/day just for verification. So the architecture became a cascade.

Most claims resolve free via fuzzy match. Only genuinely hard cases reach the LLM.