Research

Research that ships as software.

From Nagercoil, we build low-level systems for AI — GPU kernel compilers and inference engines that run anywhere a GPU runs. Open source, reproducible, and pushed to production.

Systems · Training2026

v0.15 · pip install axis-zyora

Axis — A Compiled CUDA Engine for Transformer Training & Fine-tuning

India's first proprietary AI training & fine-tuning framework. Axis lowers the entire training step — forward, backward and the AdamW update — into a single compiled C++/CUDA execution plan, and proves every result numerically identical to a reference before it ships.

TrainingFine-tuningLoRACUDAbf16Flash attentionMulti-GPUZeRO-1

Read the research GitHub

38%

Less GPU memory for LoRA vs HF + PEFT

99%

Multi-GPU weak-scaling efficiency (2×A100)

3.1e-07

Reference parity over 30 training steps

1 call

Forward + backward + AdamW per step

Systems · Inference2026

v2.0 · Open source

ZSE — Zero-dependency Server Engine for LLM Inference

A production LLM inference engine that owns the full stack — no PyTorch, no Triton, no transformers. Models load in seconds, not minutes, and serve in a fraction of the memory other engines need.

LLM servingINT4 quantizationCUDAROCmMetalRAG

Read the research GitHub

30.2×

Faster cold start vs vLLM (T4)

5.79 GB

VRAM for Qwen2.5-7B on a T4

~5 MB

Pip install · zero ML deps

40.6k

Lines of code · 0 dependencies

Compilers · GPU2026

v1.0 · Open source

locomp — A Python GPU Kernel Compiler for Apple Silicon, NVIDIA, AMD & RISC-V

Write a GPU kernel once as a plain Python function. locomp compiles it through an SSA intermediate representation into native Metal, CUDA, HIP or RISC-V vector code — one kernel runs everywhere, no rewrites.

CompilerSSA IRMetalCUDAROCmRISC-V RVV

Read the research GitHub

Hardware backends from one source

85×

Faster GELU vs PyTorch (A100)

2.6×

Faster GELU vs MLX (Apple M1)

227+

Tests across M1, A100, RISC-V

How we work

Our research principles

Own the full stack

We rebuild from first principles — kernel compilers, inference engines, formats — instead of stacking dependencies. Fewer layers, more control.

Numbers, not marketing

Every claim is reproducible on real hardware. We publish the benchmarks, the scripts and the honest limitations alongside the wins.

Open by default

Our research ships as Apache-2.0 software you can read, run and build on — from a laptop GPU to a data-center cluster.

Want to collaborate or sponsor?

We work with researchers, hardware vendors and open-source backers pushing efficient AI infrastructure forward.

Get in touch Explore our code