Hey there 👋

I'm Mahyar Jahani-nasab

Recognized (to my surprise) by peers for having a knack for research, I’m happiest when I’ve thrown myself into a brand-new scientific rabbit hole—somehow managing to learn the latest advances fast enough to come up with fresh ideas before I get completely lost. My real curiosity lies where neuro-symbolic, generative AI, physics-informed machine learning, and heavy mathematical frameworks all meet—basically, the nerd’s Bermuda Triangle. I balance theory with practice: I can wrangle hundreds of lines of code, build sturdy software pipelines, and occasionally forget what day it is while polishing a solution until it actually looks elegant.

Recent Research Endeavors </>

Assured Neuro-symbolic Sequential Decision Making with Physics Informed Neural Networks
Ph.D.

Working on making new AI systems that combine thinking in clear steps with learning from real-world physics using physics informed machine learning and neuro-symbolic.

Redesigning RNNs for Fully Physics-Informed Loss Function

Training Recurrent Neural Networks (RNNs) solely through physics-informed loss functions presents significant challenges, necessitating numerical derivatives between blocks. In this work, we introduce structural modifications to RNNs that eliminate the dependency on numerical derivatives, allowing for the formation of purely physics-informed loss functions using backpropagation. The archived version of this paper is currently accessible, with the source code set for release upon official publication.

Enhancing Convergence Speed of PINN
OSS

Streamline your training with PINNs! Dive into our newly published paper and access the source code to double the convergence speed and boost the success rate of your Physics-Informed Neural Networks.

Latest from blog

Building a Scalable Data Pipeline for LLM Training: From Streaming to Production

A deep dive into creating an enterprise-grade data collection and processing pipeline for Large Language Model training, featuring async processing, quality control, and tokenization at scale.

Closing a Chapter, Opening a New One

Reflecting on my journey as an AI Software Engineer at Solico Group and announcing my next step as an AI Researcher at Monash University.

Adapter for fine tuning LLM
OSS

Creating a patch for integrating adapter modules into the RoBERTa model allows for efficient fine-tuning tailored to specific tasks. By embedding lightweight adapters within the pre-trained layers of RoBERTa, we can freeze the majority of the original parameters, updating only the new adapter parameters. This approach significantly reduces computational overhead and speeds up training. Implementing this patch involves defining the adapter architecture, integrating it within RoBERTa, and focusing the training on these adapters to leverage pre-existing knowledge effectively. Don't forget to check the source code!

Assured Neuro-symbolic Sequential Decision Making with Physics Informed Neural Networks Ph.D.