Skip to content
pathcosmos

pathcosmos

Building AI from the Ground Up

From LLM Architecture to Deployment

View Services

Services

🧠

Custom LLM Development

Design, training, alignment, and deployment — all built from scratch.

FRANKENSTALLM, EVAFRILL-Mo

⚙️

AI Infrastructure

GPU monitoring, ML environment automation, and file processing infrastructure.

mig-gpu-mon, machine-setting, filewallball

🚀

AI Application Solutions

Video privacy, document automation, and other domain-specific solutions.

jeju_masking, hwpx-generator

Portfolio

FRANKENSTALLM

3B Korean Dense Transformer LLM, built from scratch

Korean NLU 100.0142.5 TPS8×B200 FP8 Training
TransformerGQAFP8ORPOSentencePiece

EVAFRILL-Mo

2.94B Hybrid Mamba-2 + Transformer, Nemotron-H inspired

24 SSM + 2 Attention Layers55B TokensSLERP Merge
Mamba-2Nemotron-HHybridDPOFP8

jeju_masking

YOLO12xTensorRTPrivacy

mig-gpu-mon

RustNVMLMIG

hwpx-generator

PythonHWPXCOM

machine-setting

CUDAPythonAutomation

dy_gh_filewallball

FastAPIDockerREST

About

Built from Scratch

No forks, full custom implementation. All design decisions and failures made public.

Training

Cloud GPU Gov't funding in progress

Pipeline

Pretrain → SFT → DPO/ORPO → Deployment

Transparency

Builder logs, failure records, all decisions public