Snath AI · Open Source Research Initiative

Research first.
Product second.

Snath AI is an independent research initiative with one thesis: the science of how AI systems learn continuously, without forgetting, and without human annotation — published openly before it becomes a product.

No institutional funding No lab affiliation Apache 2.0 · CC BY 4.0
6
Permanent DOIs
DAS · UCR · LTL · EIM · PAV · PERSIST
33
Named invariants
10 abstract contracts · UCR
0.94
AUROC annotation-free
from 0.45 baseline · LTL
The Lár Research Series

Six papers. One architecture.
Each one opens the gap the next one closes.

01 DAS

Divergence Is Not Noise

Zenodo
02 UCR

Universal Cognitive Routing

Zenodo
03 LTL

The Lár Training Loop

Zenodo
04 EIM

The Encoder Is Not the Memory

Zenodo
05 PAV

Physics Assumption Violations

Zenodo
06 PERSIST

PERSIST

Zenodo
The Engine

The same architecture that runs the science runs the product.

Lár is the deterministic graph execution engine the whole programme is built on. Every compliance primitive, every routing invariant, every adapter in the papers — it all runs on Lár. Pure Python. No magic loops. Every decision path is explicit code.

Architecture property
from lar import GraphExecutor
from lar.compliance import AuditLogger
# same engine, same guarantees
# whether you route a robot or
# audit a €500k loan decision
executor = GraphExecutor(graph)
result = executor.run(task)
# HMAC-signed · deterministic · resumable
Research & Engineering

Aadithya Vishnu Sajeev

Founder · Snath AI Open Source Research Initiative

Self-taught architect. Six papers in the Lár series, all authored independently. Background in computer science and cognitive systems. Currently based between Kerala and Ireland.

Product & Design

Vinay S

Co-founder · Product & Design

Product architect and designer behind the Snath AI identity, developer experience, and the compliance showcase. Shaped how the engine becomes a product users can reason about.

What we believe

Open by default

Code under Apache 2.0. Papers under CC BY 4.0. Every result deposited on Zenodo with a permanent DOI before it becomes a product claim.

Reproducible

Runs are HMAC-SHA256 audited. Seeds are published. The numbers on the research page are the numbers in the repo.

Independent

No institutional funding. No lab affiliation. All compute is personal. That constraint forces efficiency.

Falsifiable

Experiments are pre-registered. We publish honest negatives. There is a standing list of what we have not proven on the research page.

Start with the science.

Six published papers with DOIs, runnable code, and honest accounts of what isn't proven. The product is built on top of it.