Built MSOrchestrator , It is a local agentic system for materials screening. Ask in plain English...Built MSOrchestrator , It is a local agentic system for materials screening. Ask in plain English...
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
Vinay 's avatar
Vinay K Rajith
• Sep 8
Built MSOrchestrator , It is a local agentic system for materials screening. Ask in plain English, an 8B model calls tools, and graph neural networks return a crystal property in under a second. No DFT job.
I trained a CGCNN from scratch on 4,000 structures with MAE 0.920 eV. A pretrained ALIGNN I downloaded in 12 seconds with MAE 0.816 eV. Nearly identical, so I almost called it a tie.
Then I checked rank correlation, which is what a screen actually runs on: −0.014 vs +0.825. One ranks at chance. One ranks well. Average error could not tell them apart.
At scale 5,000 stable compounds where the pretrained model held at ρ = 0.742. On a 12-candidate benchmark against experimental band gaps it beat the DFT values it approximates.
Every prediction carries its theory level, provenance and caveats, and refuses outright when no trained checkpoint exists. Runs entirely on one laptop.
So the local Nemotron-8B orchestrates the tool calls and a pretrained graph neural net answers them
https://github.com/Vinay-K-Rajith/materials-orchestrator
#AppliedAI #MaterialsScience #MachineLearning #GraphNeuralNetworks #LLM #Python
Post image
Back to feed
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started