AI reliability · builder · independent research

I build systems that make AI prove itself.

I’m Reyaansh Sinha, a young developer building real systems around AI reliability. Not just experiments. Tools that check claims, catch contradictions, and ask whether a model is actually correct.

verification correctness systems
Explore Halgorithem Open GitHub
02 active builds
AI outputs under test
Real data over guesses
Active research lab mindset
about

Young developer. Independent researcher. Builder first.

I care about AI reliability because the hard part is not making a model sound impressive. It already does that. The hard part is checking if the answer survives contact with data, logic, and someone who is actually paying attention.

My work is about challenging AI, not just using it. I build experiments, tools, and prototypes that push on the places where models are confident and quietly wrong.

Some of it is rough because it is real. Build, break, tighten the system, repeat.

build small, test hard logic before vibes prove it or flag it
projects

Systems in the lab.

Halgorithem

verification engine

Halgorithem validates AI-generated claims against real data. It focuses on hallucinations, contradictions, and answers that look clean but do not hold up.

It is not a typical RAG system. The direction is more deterministic and logic-driven: claims, sources, rules, conflicts, evidence.

truth verification over generation gold standard for AI correctness

> initializing halgorithem...

> verifying claims...

> contradiction found: confidence != correctness

Splice

experimental

Splice is a dev-focused tool taking shape around systems, interfaces, and connecting moving parts without turning everything into chaos.

It is still intentionally a little mysterious. Somewhere between platform, tool, and lab bench. The interesting part is the architecture underneath.

system-level prototype-heavy still cooking
philosophy

The part I keep coming back to.

AI is powerful, but not trustworthy by default.
Most people focus on making AI smarter. I focus on making it correct.
The future is not just generation. It is verification.

> models should explain themselves in ways a system can check

> confidence is not evidence

> correctness needs a standard, not a mood